Home

Evaluating Spatial Normalization Methods for the Human Brain

1. END OF FOR LOOP brainlistlengths is a vector that lists the number of lang sites per brain in the order the brains will be in Galedist 2k matrices brainlistlengths lt c brainlistlengths row 113 HEHE HEH EH EEE EEE EE HH EEE EE EHH HE EE HE EHH EE EEE OH OE HE EE EE EE EE FE EEE EE EE Er Ep Ep Ep EE CALCULATION OF NUMBER OF LANGUAGE SITES FOR EACH BRAIN AND TOTAL NUMBER OF ROWS TO BE USED LATER HEE EE HE EH HH HE EE EEE HE EEE EE EEE EH HEE EH EHH EEE HE EE EE EH EE EE EE EEE EOE EE EE EEE HH EE H length x length lang prenorm list totalLength U RowNum vec lt vector length 0 for k in 1 1length tempRowNum lt nrow lang prenorm list k tempLength tempRowNum if length RowNum vec gt 1 tempRowNum c RowNum vec tempRowNum J RowNum vec lt tempRowNum totalLength lt totalLength tempLength END OF FOR LOOP itt HH Ht Ht HH HH HH HH d EH HOE EH HOE EE OEE EE EEE EOE EEE du PRENORM DISTANCES BETWEEN LANGUAGE SITES W IN EACH BRAIN itt HH HH HH HH HHH HT d EH HOE EH HOE EE OEE ESE ESE EE SEE EE T prenorm lang list lt lapply lang prenorm list functlonted i dirstaprenorm amabrzx c cbindg tcodoxcdoY cQem dist dist prenorm matrix method euclidean ttt HH HE EE HE HE HH EH HE EE EE TE EE HH EE EE EH HH EE E E EE EE EH E E EEE CARET DISTANCES BETWEEN LANGUAGE SITES w in EACH BRAIN ttt HHH HE EH HE EH HE HH EEE EE TE EE HE EE EE EE HE EEE EE EE EE HH E EEE H
2. Figure 26 Summary of error rate by CPS parcel descending left to right from the parcel with the most sites MSTG to the parcel with the least sites TrIFG 51 Error Type Analysis A paired t test of all error types showed that the methods differed on average by less than a tenth of an error per subject The confidence interval 1 2 1 4 revealed that either method could be better than the other by more than one error per subject Analysis of type 3 errors revealed the average difference for this error type was less than one error per subject Caret normalization could result in as much as two errors less per subject than SPM2 while a SPM2 normalization could result in not more than 1 error less per subject than Caret Type 2 error differences were statistically significant p lt 01 The confidence interval showed that a SPM2 normalization could result in more than 2 errors less per subject than Caret while a Caret normalization could result in not more than error less per subject than SPM2 Type 2 errors are discussed in Section 5 Error Type Analysis number of errors 20 30 10 Error 1 Error 2 Error 3 Figure 27 Break down of same type and unique mapping errors Orange bar indicates errors unique to Caret Blue bar indicates errors unique to SPM2 Black bar indicates errors shared by both methods 52 Same Error Type Mappings Figure 27 breaks down errors by type and whether the error was unique to o
3. ee m areg 4 temprow lt post dat 1 temp matrix lt rbind temprow langcoord caret langcoord caret lt temp matrix END OF FOR LOOP 112 itt HH HH HH HHH HH HH OE HOH EE H SPM2 LANGUAGE SITE COORD itt HH HH HH HHH HH HH OH EE HEE H Tandgeooroewspm lt Dost dat Los temp matrix lt langcoord spni Length dam post dat 1 Porc mm L lbengcth 4 PEC POSE Oat et CoM Reg ron seb lt 6 postydeatl i Aloorrcham ee USpM25 Ar 4 temprow Ee Da 0 temp matrix lt rbind temprow langcoord spm langcoord spm lt temp matrix END OF FOR LOOP tae aE dg dd Ae a ae OE AE ae aE aE AEE aE EEE aE aaa EEE EEE EEE EET T CREATE LIST OF LANGUAGE SITES SEPARATED BY BRAIN ID aE aE dg dd Ae a ae OE TOTO dd EEE d dd aaa lang prenorm list split langcoord prenorm langcoord prenorm Brain ID vd dg dd dg OE AE a aE EAE aE EEE d dd dg dd dg dd dg dd ATTEMPT TO GET BRAIN LANG SITES IN CORRECT ORDER tae at dg dd Ae a OE AEE AEE AEE aE Ea EAE d dd dg dd EET brodndace lengecoord orenorn ee De en DE MAI length lt dim brainlist 1 brainID 0 row U tempNumRow U brainlistlengths vector length 0 EE Las Lengel 4 if row 0 brainID lt brainlist m Brain ID row 1 else if brainlist m Brain ID brainID row row 1 else brainlistlengths c brainlistlengths row row 1 brainID lt brainlist m Brain ID
4. 11 4 13 Select AC PC Aligned 11 4 14 Select Done 11 4 15 exit AFNI 11 4 16 The result of this process is two files Pxxx Exxxxx Sx 111 acpc HEAD and Pxx Exxxxx Sx_111 acpc BRIK which will be used in the next steps 11 5 Vecwarp preparation 11 5 1 Go to Pxxx SURFACES directory 11 5 2 Copy Pxxx xxxxx Sx_111 L full segment_vent_corrx fiducial magctr xxxxx coord to Pxxx directory 11 5 3 Create in vec file as follows 11 5 3 1 At command line mincinfo Pxxx xxxxx Sx 111 mnc which will give you the x y z coordinates of the magnet center of this volume Note that results of mincinfo list the coordinates z y x 11 5 3 2 In a text editor create the in vec file with this data 1 0 0 x H d y U 0 1 z 101 11 Create CSM coordinate files continued 11 5 4 Create out vec file as follows 11 5 4 1 Call 3dinfo Pxxx Exxxxx Sx _111 acpc HEAD The output will provide three rows starting with R to L extent A to P extent I to S extent In the first row we want the value associated with L In the second row we want the value associated with P In the third row we want the value associated with I 11 5 4 2 In a text editor create the out vec file with this data 1 U U 11 6 The in vec and out vec files will be used to translate the coord file to full brain grid for input to the ACPC warp The out vec file will be used to translated the ACPC warped coord file to AC center Run Vecwarp on the coord file first at the comman
5. 2 2 8 Subject Name a unique identifier for each individual brain Do not include the hemisphere in the subject name Typically use the default name provided Investigator Name of individual responsible for the study or segmentation Group UW SIG Data Type MRI Resolution 1 0 mm Species human Comments PXX identifier Volume Extent SureFit preserves information about the position of each cropped volume within the original image volume This is useful when aligning structural and functional MR data from the same subject e Volume Already Cropped Typically the original volume will be uncropped thus select no e Hemisphere Typically select both LR e Region Typically select entire cerebral hemisphere one or both e Filename select change and accept default file name 2 3 Select Volume Orientation E SureFit s conventions are LPI which means the left hemisphere is displayed on the left side Typically volumes loaded from our files will be in the correct orientation Follow the instructions in the Volume Oriention window to ensure that the volume is indeed correctly oriented If the volume is correctly orientated select yes If the volume is not correctly oriented select no and follow the steps outlined to get the correct orientation and polarity Once this has been achieved select Save This will create a volume in which Orient mnc is appended to the initial volume name and use this as Voll for subsequent proces
6. 9 2004 there has been a clarification on the starting point of the lateral fissure landmark border that we believe will impact the normalization in the middle part of the superior temporal gyrus and sulcus STS STG the region where there is important variation between the subjects and colin27 s cortical folding patterns and the region to which most errors were attributed Specifically redrawing the lateral fissure landmark border for each of the 11 subjects according to the clarified guidelines is expected to constrain the STS STG more medial dorsally which will tend to reduce differences summarized in Figures 32 and 33 Given this landmark revision we would expect to see a marked decrease in type 2 errors as a result of Caret normalization Cost Benefit Analysis Using SPM2 to normalize the CSM coordinates required notably less user input and time than using Caret because SPM2 does not require a surface reconstruction for normalization to the target To run SPM2 normalization the only input required is a Minc file of the MRI and a text file including the original CSM coordinates With this input an experienced analyst can generate 63 the normalized cortical site coordinates in approximately 20 minutes or less Using the surface base method the input required includes a surface reconstruction which requires a total of 1 hour of interactive processing and 1 2 hours for segmentation Once the surface reconstruction 1s complete th
7. Copyright 2005 Veronica Susanne Smith Evaluating Spatial Normalization Methods for the Human Brain Veronica Susanne Smith A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Electrical Engineering University of Washington 2005 Program Authorized to Offer Degree Department of Electrical Engineering Abstract Evaluating Spatial Normalization Methods for the Human Brain Veronica Susanne Smith Chair of the Supervisory Committee Professor Linda S Shapiro Department of Electrical Engineering Department of Computer Science Cortical stimulation mapping CSM studies have shown cortical locations for language function are highly variable from one subject to the next Because no two cortical surfaces are alike and language is a higher order cognitive function observed variability is attributable to a combination of functional and anatomical variation If individual variation can be normalized patterns of language organization may emerge that were heretofore hidden In order to discover whether or not such patterns exist computer aided spatial normalization is required Because CSM data has been collected on the cortical surface we believe that a surface based normalization method will provide more accurate results than will a volume based method To investigate this hypothesis we evaluate a surface based Caret and volume based method SPM2 For our application the ideal met
8. Note the landmark traces that correspond to the inflated surface traces in Figure 13 39 Volume based normalization To normalize the source to the target using SPM2 we input the subject s MR image in the form of a Minc file with X increasing from patient left to right No flipping was done during normalization With the exception of using the template bounding box and cubic 1 mm voxel dimensions the default spatial normalization settings were used The selected template image was a T1 Minc average volume of the MNI152 discussed in Section 2 3 Header and image warp files were automatically written Then the Deformation function which writes header and image deformation files using the normalized mat file as input was called Next the Invert Deformation function which writes header and image inverse deformation files based on input of the individual s Minc volume was called The interactive normalization process took 15 20 minutes per subject on a Dell Dimension dual 450 Mhz processor running Debian Linux 40 Step 6 Evaluation Language Site Spread Reduction In order to measure the change in distances between language sites across subjects prior to and after applying selected spatial normalization methods we needed to first measure these distances The following equations show how we measured distances between any two subjects We expanded these calculations for each of the 11 subjects and calculated 4B PR RER B p2 p
9. Post norm flat map with normalized cortical sites on colin27 on the right The language site green has dropped into the sulcus P164 CSM sites visualized on P164 flat map P164 normalized CSM sites visualized on colin27 flat map Figure 20 Example of Type 3 Error P164 pre normalized location of motor site red circled in blue is mapped incorrectly as seen on the right The post normalization location has been moved across a sulcus from VPrG to VPoG 45 46 Section 4 Results To determine which spatial normalization method would best transform cortical stimulation mapping data to a target atlas we measured both spread between statistically significant language sites across subjects and preservation of anatomical localization as described in Section 3 Table 6 summarizes the results of the analysis for both measures In 2D space Caret reduced the average distance between language sites across subjects 3 mm more than SPM2 Caret reduced the average 3D distance between language sites across subjects 6 mm more than SPM2 Using the jackknife estimate of variance we did not show a statistically significant difference between the surface based and volume based methods Efron and Tibshirani 1993 Table 6 Summary of Results 2D Spread Reduction 3D Spread Reduction Localization Accuracy Rate 6 9 mm 2 8 mm 79 6 3 9 mm 2 2 mm 78 0 Anatomical localization accuracy rates as analyzed using a paired t test revealed a statis
10. for both methods the resulting accuracy rate for both methods would increase by more than 9 resulting in 88 89 accuracy for Caret and 87 37 accuracy for SPM2 Discussion on how to improve mapping accuracy follows in Section 5 Unique Error Type Mappings Unique Caret errors 16 accounted for 25 of all Caret errors Analysis of these errors reveals that 31 of the errors should have been mapped to the middle part of the superior temporal gyrus MSTG The other two parcels with the most common errors were the posterior part of the superior temporal gyrus PSTG with 19 of errors and the middle part of the middle temporal gyrus MMTG with 13 of errors Type 2 errors accounted for 56 of unique Caret errors Type 3 errors accounted for 38 of these errors and there was a single type 1 error 6 Table 10 Summary of errors unique to the Caret spatial normalization method P54 21 MSTG 0 5 2 sulcus PSTG 0 5 2 sulcus ASTG 1 3 no AMTG P60 32 PMTG 0 5 2 sulcus P60 40 PSTG 13 nofPMTG Pol 22 ASTG 05 2 sulcus MSTG 0 5 2 SUUS AMTG 1 3 net MITG MMTG 1 no PSTG MPoG 0 5 2 sulcus P117 32 PSMG 1 no PMTG P170 26 MSTG 0 5 2 sulcus 54 Table 11 Summary of errors unique to the SPM2 spatial normalization method SPM2 P54 VPrG no VPoG P55 AnG no PMTG P58 PSTG PMTG P58 ASTG no VPoG P60 MITG no MMTG Pol MSTG no PMTG P62 VPrG 0 2
11. supramarginal gyrus SMG and terminal ascending segment of the lateral fissure figure 30 These uncommon localized folding patterns of the colin27 hemisphere help explain the average error rates of 50 or more in the VPrG and PSMG circled in figure 30 Figure 30 The colin27 atlas lateral left hemisphere surface reconstruction with areas of uncommon cortical folding patterns circled in red 59 Parcels in this area contain 175 of 198 mapped CSM sites Figure 31 Most common parcels for mapped CSM sites are circled in red with the average error rate listed for each parcel Our analysis comparing a digital atlas of 12 normal subjects PALS B12 to 10 of our 11 epileptic subjects revealed that epileptic subjects have a broader superior temporal gyrus STG than the normal subjects figure 32 Analysis of a sulcal depth difference flat map revealed that the greatest difference between epileptic and normal subjects left hemispheres 1s in the middle part of the superior temporal gyrus MSTG on the CPS scheme figure 33 60 average epileptic MSTG shape average normal MSTG shape Figure 32 Comparison of inflated left hemisphere average surface reconstruction of 10 epileptic subjects in this study left to 12 young adult normal subjects included in the PALS B12 atlas right Region of most striking differences in sulcal depth This region corresponds to the MSTG parcel on CPS Figure 33 Sulcal difference m
12. 0 300000 0 600000 0 500000 0 500000 300 20 smoothing parameters 1 1 000000 100 20 10 5 morphing parameters 1 1 0 300000 0 600000 0 500000 0 500000 300 5 smoothing parameters 2 1 000000 50 20 10 1 morphing parameters 2 1 0 300000 0 600000 0 500000 0 500000 300 2 flat parameters 900 0 000010 1 000000 20 target directory target spec Human POP AVG L REGISTER Normal B6 with INDIVIDUAL 73730 spec target landmark border Human POP AVG Normal B6 Projected L SPHERE border target closed topo POP AVERAGE L Human sphere 6 73730 topo target cut topo POP AVERAGE L Human sphere 6 73730 topo target sphere coord POP AVERAGE L Human sphere 6 73730 coord target fiducial coord DOP AVERAGE L Human sphere 6 forFIDUCIAL 73730 coord target flat coord output spec file deformed 4K NoFidHuman colin L REGISTER with POP ATLAS 73730 spec sphere fiducial sphere ratio false 0 500000 inverse deformation false DATA START
13. 13 19 File Save Window as Image spm Pxxx flat jpg 13 20 Switch to inflated fiducial views and save those captures if desired 13 21 Save flat foci coords as follows 13 21 1 File Save Data File Foci File 13 21 2 Foci Associated with Surface Type Flat 13 21 3 Filename spm Pxxx CSM flat foci 13 22 The spm Pxxx CSM foci and spm Pxxx CSM flat foci files will be used for for input into PostNorm csv and FlatPostNorm csv files see Appendix F 107 Appendix B Resampling MRI to 1mm Cubic Voxels Excerpt from Resample PatientBrainData xls spreadsheet BrainID P55 P58 P60 dim name xspace yspace zspace dim in voxdim in voxdim out Start dim out xspace yspace zspace dim in voxdim in voxdim out Start dim out xspace yspace zspace dim in voxdim in voxdim out Start dim out length 256 250 256 256 0 896471 l 108 1 101 123 3 229 256 256 256 256 0 86549 106 1 95 1 110 2 222 256 256 256 256 0 884706 l 111 5 113 5 115 6 226 step start 0 896471 108 1 0 896471 101 0 896471 123 3 0 86549 106 1 0 86549 95 0 86549 110 2 0 884706 111 5 0 884706 113 5 0 884706 115 6 108 Appendix C Creating a stripped coordinate file for use in SPM2 Use text editor to create file with coordinate data from the CSM database DAZU G L AII El L2 49 21S 21323 9439 04 590 md Z0 20 6 34306 056043509 761 71 59 99 Dao qud 3 9
14. 2 8 13 3 8 13 4 8 13 5 Switch to the flat map in the main window Window Viewing Window 2 set to inflated surface Toolbar L in Window 2 for lateral view of inflated surface Toolbar D C toggle on borders Select Borders from the D C menu 8 13 5 1 Toggle on Show first link red 8 13 5 2 Draw Borders as Points and Lines 8 14 Calcarine and Medial Wall borders These were drawn at flattening but generally some border points are nibbled off Note that there are distinct gaps between the medial wall dorsal and ventral segments in the colin atlas borders that are used as the landmark reference Replicating these gaps as closely as possible on the individual surfaces will reduce the probability of crossovers along the medial wall ventral segment during registration 8 14 1 Touch up of these borders as needed is done using the Layers Delete Border Point with Mouse feature in preparation for registration For more detail see Spherical Registration to Atlas Core6 landmark set link at http pub download brainmap wustl edu pub donna WASHINGTON 200503 p117 html 8 15 Identify Central Sulcus extent as follows 8 15 1 If the individual s inflated surface doesn t align to the same coronal axis as the atlas then select Z from the Toolbar s drop down menu to switch the rotation axis to Z and rotate the surface until it is roughly AC PC aligned 1 e aligned along the atlas coronal axis 8 15 2
15. 8 15 3 8 15 4 8 15 5 In PXXX identify the central sulcus on the inflated lateral view using the atlas as a guide Click on a node in PXXX s central sulcus about where the ventral tip of the central sulcus border on colin about 15mm above the edge of the Sylvian fissure where there is no ambiguity as to whether you are in the sulcus proper Click on a node along the edge of the Sylvian fissure just below the node clicked above so you can read out the distance measurement in the Identify window If the distance is 12 18mm then you re in the right ballpark Click on a node on the dorsal tip of the central sulcus Switch to dorsal view in both sessions 90 8 Drawings borders in Caret for surface normalization continued 8 15 6 The node just clicked identifies which sulcus is the right one Click on a node in that sulcus about 15mm from the edge of the medial wall again where there is no ambiguity as to whether you are in the sulcus proper 8 15 7 Click on a node along the medial wall just across from the last node clicked to make sure the distance is about 15mm from the medial wall 8 16 Identify Sylvian Fissure extent as follows 8 16 1 The Sylvian fissure landmark border begins about 12mm along its primary fundus SF on the flat map posterior to its intersection with the main secondary fundus SF2 On the inflated map this is just before the beginning of the dorsal ascending ramus of the Sylvian fissure and appears sl
16. Breslau Cohn Weigert Wesson KA Brain Basics for the Teaching Professional Science Master lt http www sciencemaster com columns wesson wesson_part_01 php gt Accessed Dec 2003 Woods RP Grafton ST Watson JDG Sicotte NL Mazziota JC Automated Image Registration II Intersubject validation of linear and nonlinear models Journal of Computer Assisted Tomography 1999 Jan Feb 22 1 153 165 72 Appendix A Evaluation Protocol 1 Patient MRI orientation and preparation 1 1 Create directory for Pxxx KS 1 3 Download ExxxxxSx mnc file from usr local dataX brainproject patients directory into PXXX directory Typically the directory will include 3 ExxxxxSx mnc for any given patient Sx assumes numerical values eg S1 S2 S3 Typically in this example S1 will be the structural MR S2 will include veins and S3 will include arteries If not confident of the content of the 3 minc files download all three and view in SureFit to confirm which is the MR file needed for normalization Resample volume to 1mm cubic voxels as follows 1 3 1 Verify volume is in correct orientation LPI by calling mincheader Pxxx Exxxxx Sx mnc at command line Typically the volume will be oriented correctly but will not have cubic 1mm voxels 1 3 2 Before resampling call mincinfo Pxx Exxxxx Sx mnc and get output like the following dimension name length step start Zspace 256 0 892941 107 6 yspace 256 0 892941 114 5 xspace 256 0 89
17. EH HH HH HE EHH EEE HE EH EEE EE EH HE EE E E EE EE HH EEE DISTANCE ACROSS BRAINS FUNCTION DAB tt HHH HE EH HE EH HH HH EE EE HH EE EE EH HE EE EEE EE EE E E E EEE HE DAB lt function brainlDs distmatrix index lt outer brainIDs brainIDs lt DAB vec lt distmatrix index DAB vec as vector DAB vec DAB vec l 119 HEE EHH HH HH HEE EHH EH HEE EE HE HE EE EH HE EE EEE EE HE HE EE EE HE HE EEE AVG PRENORM DISTANCES FOR LANGUAGE SITES ACROSS ALL BRAINS HEE EHH HH HH EE EHH EH HEE EH EHH EE EH OH EE EE EE EH EE EE EEE EE EE EE EEE PRENORM DATA SET OF LANGUAGE SITE COORD FOR ALL BRAINS QCODEPOd prenormomeLrix lt lengooord prenorm ert Tye SAS PRENORM DATA SET OF DISTANCES BETWEEN LANG SITES IN w in and ACROSS BRAINS calcdist prenorm lt dist coord prenorm matrix method euclidean Ca leas prenorm meatrix as Malrix caledist prenorm prenorm bIDs langcoord prenorm Brain ID prenorm DAB DAB prenorm bIDs calcdist prenorm matrix prenorm avgDAB mean prenorm DAB expect postnorm avgDAB 1 440741874 3333 prenorm avgDAB TETTE OE d dg dd aE OE OTT TO dd dg dd dg OE d dg dd EET ETE AVG CARET DISTANCES FOR LANGUAGE SITES ACROSS ALL BRAINS it aE aE dg dd Ae a AE Ae AE aE a aE Ea aE AE EEE aaa EEE EEE EEE EEE T Goord caret mabrix lt langooord cgretjoe XN TY TZI 4 POSTNORM CARET DATA SET OF LANG SITE COORD FOR BOTH BRAINS calcdist caret lt dist coord
18. Pxxx Exxxxx Sx_111 L full segment vent corrx xxxxx topo FIDUCIALcoord file Pxxx Exxxxx Sx_111 L full segment vent corrx fiducial magctr xxxxx coord 6 5 2 Start Caret and load all files in the just created spec file 6 5 3 File Import File Minc Pxxx LR full SMRI mnc 6 5 4 Switch to view Volume press D C and select Overlay Underlay Volume 6 5 5 Toggle show surface box in lower right ON 6 5 6 Switch between Coronal Horizontal and Parasagittal views Scroll up and down to confirm alignment of surface outline with the volume 6 5 7 If out of alignment back track to determine cause of mis alignment and redo until surface and volume alignment are confirmed 87 7 Preparation for Caret normalization Tb In the terminal shell run the following commands in the Pxxx SURFACES subdirectory cp Pxxx Exxxxx Sx lll L full segment vent corrx Surface xxxxxx spec Pxxx Exxxxx Sx III L REG with Colin Core xxxxx spec Note Before running the next commands make sure there are no spaces on either side of the characters or else you ll lose everything rm CYCLE coord Compression HighSmooth RGB paint coords as border border debug flat morph distortion surface shape spherical morph distortion surface shape TEMPLATE CUTS rm segment rm segment vent rm segment vent corr This is a good time to remove some of the intermediate patched segmentation volumes in Pxxx SEGMENTATION For example if the final segmentati
19. Type 2 Errors A paired t test of type 2 errors did reveal a statistically significant difference p lt 01 in the methods SPM2 mappings resulted in only one type 2 error compared to 18 type 2 errors mapped using Caret The average difference between methods was 1 55 errors per subject We believe that this difference is attributable to the underlying differences in normalization approaches used by the different methods The SPM2 algorithm maximizes a voxel intensity match between source and target As a result the volume based method will very rarely end up with an alignment resulting in a pre normalized gyral location e CSM sites are always on the gyrus being relocated into a sulcus where the voxel intensity is markedly less than the intensity found on a gyrus The surface based method however maximizes alignment of a set of landmarks based on cortical folding patterns without consideration for voxel intensity If the selected landmarks vary enough between the source and target then the normalized sulci and gyri will be deformed in ways that confound mapping of functional data to corresponding regions of the anatomical substrate The Core6 landmark protocol was designed to minimize this problem by selecting the most stable landmarks and constraining the extent of each landmark to regions where it is reasonable to expect good correspondence across nearly all subjects Van Essen 2005 Since performing the spatial normalization for our 11 subjects
20. Woods method located about 40 of differences versus 23 for Talairach again demonstrating the superior accuracy of non affine over affine spatial normalization methods Grachev et al 1998 Davatzikos et al compared two non linear methods SPM Statistical Parametric Mapping the most widely used method for analysis of functional activation images and STAR Spatial Transformation Algorithm for Registration They found that STAR resulted in relatively better registration Davitzikos et al 2001 SPM employs a volume based approach that minimizes the sum of the squared differences between the source image and target image while maximizing the prior probability of the transformation The maximum a posteriori solution is found iteratively the algorithm starts with an initial parameter estimate and searches from there The SPM algorithm stops when the weighted sum of square differences no longer decreases or after a finite number of iterations Salmond er al 2002 The STAR algorithm differs from the SPM approach in that it employs an elastic instead of a parametric transformation thus it has thousands of DOF compared to the relatively low DOF allowed for by SPM Additionally STAR applies surface based curvature matching along the cortex thus incorporating shape information in the matching mechanism These differences were attributed to STAR s improved registration Davitzikos et al 2001 SPM is one of many volume based non linear spat
21. bP Hoye EE 55 I5 Refined Spread As tee TT ose PR ER an DO en ed b aine 57 14 Correcthy Mapped Laneu Ee 57 ACKNOWLEDGMENTS Thank you to Donna Hanlon and David Van Essen for their invaluable collaboration A special thank you to Richard Martin for being a great mentor and key contributor To Dr George Ojemann I offer sincere appreciation for your years of work in neuroscience and suggesting this thesis topic Thanks to Anthony Rossini and Thomas Lumley for biostatistics consulting Many thanks to Linda Shapiro Jim Brinkley David Corina and the University of Washington Structural Informatics Group and Foundational Model of Anatomy colleagues for your ongoing encouragement feedback and support I am forever grateful for the encouragement support and guidance I received from Amy Feldman Bawarshi and Frankye Jones from my beginning efforts to the attainment of this degree To Jennifer Girard Ron Howell and Craig Duncan I offer thanks for putting your support into words and believing that I could change tracks successfully Hank Hayes and Abigail Van Slyck were also supportive mentors long before as well as during this effort thank you To Korin and Lois Haight I offer my gratitude for being my most precious chosen family Chris Galvin was there especially during the tough times to help keep perspective and light the way when I couldn t see where I was going To my mother I offer my thanks for being more than a survivor and showing me
22. bit more visible and click on nodes along the calcarine sulcus and medial wall as shown in the calc medial jpg 5 16 Select Layers Border Draw Border 5 17 Select MEDIAL WALL as the border name 5 18 Press Apply and draw a border in the main caret window on the compressed medial wall view of the spherical surface Follow the route traced by the green ID nodes for the medial wall Shift click to complete the border 5 19 Select CalcarineCut as the border name 5 20 Press Apply and draw a border in the main caret window on the compressed medial wall view of the spherical surface Follow the route traced by the green ID nodes for the calcarine sulcus Make sure the calcarine border crosses the medial wall border Shift click to complete the border 83 5 Surface Flattening continued 5 21 Select Layers Border Delete border with mouse and select the template borders for the medial wall and calcarine not the ones you just drew but the template borders the redrawn ones replace 5 22 If necessary delete existing Sylvian cut and redraw cut so that it does not cross Superior Temporal Gyrus since this will be a registration landmark 5 23 Click Continue Flattening on the dialog you pulled to the side 5 24 After a while an Initial Flattening dialog gives you the opportunity to make cuts if there are no visible red patches of crossovers click continue flattening and accept the default parameters on the following two dialogs Flat an
23. different modalities and also the registration of images with a given coordinate system The term normalization is usually restricted to the intersubject registration situation and is the term we will use in this paper Spatial normalization accuracy is a critical step to accurate quantitative analysis of the human cortex and is the focus of this research Normalization is a form of alignment that involves two parts 1 Positional normalization transformation determination of a transformation that relates the position of features in one image or coordinate space to the position of the corresponding feature in another image or coordinate space The symbol T will represent this type of transformation 2 Intensity normalization transformation determination of a transformation that both relates the position of corresponding features and enables us to compare the intensity values at those positions The symbol T will represent this type of transformation Using the language of geometry we refer to the normalization transformation as a mapping Hill et al 2000 The problem of accurately mapping data across subjects 1s confounded by two factors anatomical variation and functional variation 1 2 Anatomical Variation Notorious for the irregularity in depth and patterning of cerebral cortex convolutions the human brain structure varies notably from one person to the next The human brain is an organ that is exponentially more complex t
24. methods required that we explicitly select tools for surface reconstruction surface flattening a target atlas and spatial normalization 2 1 Surface reconstruction tools There have been many efforts to develop automated and semi automated methods for reconstructing the convolutions of the cerebral cortex The tools surveyed are available outside the laboratories in which they were developed Surface Reconstruction by Filtering and Intensity Transformations called SureFit was designed at the Washington University Van Essen lab and is based on an underlying physical model of cerebral cortex and its appearance in structural MRI The cerebral neocortex consists of a slab like sheet of gray matter with approximately uniform thickness folded into outward folds called gyri and inward folds called su ci The transition from gray matter to the underlying white matter is called the inner boundary The cortical surface called pial is where the gray matter meets the cerebrospinal fluid CSF and defines the outer boundary Van Essen et al 2001 Gyri outward folds b onn 5 ES Rr n weer ON ww En Outer gray CSF pial boundary oe Mid cortical Radial axis E thickness inner to outer WR E Inner WE gray white boundary Sulcus inward fold Figure 5 A schematic model showing a patch of folding cortex 18 SureFit generates a set of probabilistic maps for the location of gray matter whi
25. moving forward 11 8 Create AC PC aligned coordinate file as follows 11 8 1 Copy the Pxxx CSM foci to the Pxxx directory 11 8 2 In the Pxxx directory run the Vecwarp protocol on the foci file as follows 11 8 2 1 Vecwarp matvec in vec input Pxxx CSM foci gt pxxx CSM fbg coord 11 8 2 2 Vecwarp apar Pxxx Exxxxx Sx_111 acpc HEAD input pxxx CSM fbg coord gt pxxx CSM warp coord 11 8 2 3 Vecwarp matvec out vec input pxxx CSM warp coord gt pxxx CSM acpc coord 11 8 3 The pxxx CSM acpc coord is used for the pre normalized coordinates input into to the algorithm for measuring spread reduction metric See Appendix F for the source code and input csv files 103 12 Caret normalization of CSM coordinates to atlas coordinates 12 1 Open Caret5 in COLIN L LANDMARKS REG with INDIVIDUAL COREO directory 12 2 Select caret P117 E10043 S4 111 L REG with Colin Core6 71785 spec and make sure these files are selected but don t Load files yet CLOSEDtopo file Human colin Cerebral L CLOSED 71785 topo CUTtopo file Human colin Cerebral L CUTS 71785 topo FIDUCIALcoord file Human colin Cerebral L FIDUCIAL SPM2 03 12 71785 coord INFLATEDcoord file Human colin Cerebral L INFLATED 71785 coord FLATcoord file Human colin Cerebral L FLAT CartSTD 71785 coord foci file caret P117 MAG 71785 foci if not an option see below to apply deformation foci color file CSM2 focicolor surface shape Human colin Cerebral L 71785 surface shape 12 3 Make su
26. p value 2 pnorm abs mean differences se differences lower tail FALSE Input Files Excerpt from FlatPostNorm csv Brain ID CSM Region Site Number Input Files 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 54 BR BR N A BR WN e HB HR BH WO W WW WW WO WO NN A ka Ne N d E NB E HR t Ga HR HBR BR WO WO WO W WO Lu 10 12 13 14 15 16 18 19 20 21 22 23 24 25 26 30 3l 35 36 9T 42 43 10 12 13 14 15 16 18 19 20 21 22 23 24 25 26 X 11 18795 6 763459 35 895584 35 293655 5 266497 38 877956 82 049553 46 575829 38 191551 90 022171 85 187706 10 156957 12 421206 7 401104 41 34903 39 649734 10 001882 92 752426 83 236671 78 551491 104 180176 22 400347 116 294731 75 691322 17 78986 2 366655 25 799614 27 175005 2 620024 6 860496 45 04126 43 134949 40 237881 75 08091 70 536285 12 022094 8 731981 3 425389 40 967079 44 068794 8 862825 45 548531 43 534492 57 411488 82 270172 78 978302 11 708981 64 932083 33 86076 23 418528 6 319094 2 844942 0 884871 1 123019 47 406235 31 792488 61 294136 19 398674 19 405651 10 833484 39 605854 4 385891 10 695324 36 96801 13 352321 55 623337 47 554543 59 168766 85 125488 91 754105 18 850975 68 490898 41 196213 23 137539 7 097998 9 477371 3 458286 8 311852 48 751202 37 846
27. points on each source and target landmark contour The landmarks are then used as constraints for the deformation algorithm The deformation entails using Laplacian differential operators constrained to the tangent space of the sphere and basis functions that are expressed as spherical harmonics FreeSurfer http surfer nmr mgh harvard edu FreeSurfer is a software suite developed by Anders Dale and Bruce Fischl at Massachusetts General Hospital s Martinos Center for Biomedical Imaging and CorTechs Lab Inc Freesurfer employs a spherical transformation to establish a uniform surface based coordinate system Using this coordinate system points on any of the surface representations for a given subject can be indexed Freesurfer employs a procedure that aligns a cortical hemisphere with an average surface based on an average convexity measure By maximizing the correlation of the convexity measure between the individual and the average the procedure computes an optimal mapping to a 26 canonical target Fischl et al 1998 The FreeSurfer algorithm is very similar to Caret except that FreeSurfer normalization uses all sulci in maximizing correlation instead of a selected set of landmarks as is the case for the Caret algorithm There is some evidence that the limited landmark method may be superior but more evidence is needed to exhaustively compare these registration methods Desai 2004 BrainVoyager http brainvoyager de Brain
28. that I could make my own way in the world To my father I offer my sincere appreciation for helping me excel academically To Natalie I offer my deepest gratitude for making this opportunity possible and for your unwavering love and support throughout this journey vi DEDICATION In honor of Frederick D Hamrick III and Carolyn McGee Hamrick Section 1 Introduction 1 1 Registration of Medical Images Within medical research and especially in neuroscience medical images are used to investigate diagnose and treat disease processes as well as understand normal development In neuroscience research studies it 19 often desirable to compare functional and structural images obtained from the brains of patient cohorts In addition the amount of data produced using ever improving technology for generating medical images increases exponentially with each successive generation of imaging systems It is essential therefore to have reliable efficient and accurate methods for comparing and combining structural and functional brain images across subjects While the problem of comparing the brains of different individuals is an old one the development of computer aided alignment referred to as intersubject registration or spatial normalization has been substantial in the last decade We use the term registration to mean determining the spatial alignment between images of the same or different subjects acquired with the same or
29. that incorporate aspects of both volumetric transformations and surface based matching They include hybrid volumetric and surface warping Liu et al 2004 and hierarchical attribute matching mechanism for elastic registration HAMMER Shen and Davatzikos 2002 diffusing models Thirion 1998 and robust multi grid elastic registration ROMEO Hellier and Barillot 2003 Researchers want to know What are reasonable expectations for each registration method Crum has observed that there is a problem in the neuroimaging community in that we do not usually know the quality of non linear registration methods We lack the necessary framework to explicitly estimate and localize error for non linear registration tools He argues that as research studies become more sophisticated it is increasingly important to understand and measure the degree regional variation and confidence in the correspondences established by any given registration The solution lies in measuring quality at all stages of a non linear registration task We must prescribe success criterion quantify 1 technical image quality 11 relationship quality between underlying biology and imaged features and 111 registration quality Crum et al 2003 Hill concurs that it is desirable to determine both the expected accuracy of a technique and also the degree of accuracy obtained for any given set of data Hill et al 2001 11 Like Hellier and Crivello we wanted to evaluate
30. the differences between spatial normalization procedures by testing them on a given neuroimaging data set to determine which method should be chosen for analysis Hellier et al 2003 Crivello et al 2002 Other studies comparing normalization methods in the literature include Fischl et al 1997 Gee et al 1997 Grachev et al 1999 Minoshima et al 1994 Roland et al 1997 Senda et al 1998 Sugiura et al 1999 Currently the Neuroimaging Visualization and Data Analysis lab NeuroVia at the University of Minnesota is conducting an evaluation of several spatial normalization methods including AIR SPM ANIMAL HAMMER PASHA ROMEO and DEMONS http www neurovia umn edu neurovia html These studies used a variety of evaluation metrics including the dispersion metric of selected landmarks differential characteristics tissue classification spatial homogeneity of selected anatomical features such as major sulci overlap percentage of restricted volume of interest cross comparison of 3D probability maps and visual assessment Given the data set collected from cortical stimulation mapping for language localization we wanted to know what would be the best spatial normalization method to use for intersubject registration to a canonical brain atlas Friston states that goodness of a spatial transformation can be framed in terms of face validity established by demonstrating the transformation does what it is supposed to construct validit
31. the main SPM2 main menu becomes active again 10 10 Select Deformations Invert deformation 10 11 Select deformation field y Pxxx Exxxxx Sx 111 img 10 12 Select image to base inverse on Change filter to read mnc instead of mg and select Pxxx Exxxxx Sx Ill mme 10 12 1 Output files will be generated a few minutes later when the SPM2 main menu become active again iy Pxxx Exxxxx Sx III hdr and iy Pxxx Exxxxx Sx lllamgandiy Pxxx Exxxxx Sx lll mat 10 13 Quite SPM2 98 11 Create CSM coordinate files 11 1 Create coordinate file from CSM database 11 1 1 Import coordinates from CSM database by copying and pasting magnet coordinates into a text editor and adding header and assigning site types 11 1 2 Coordinate file format BeginHeader comment date Fri Feb 04 2005 encoding ASCII EndHeader tag version 1 tag number of cells 4 tag number of comments U tag BEGIN DATA 0 62 83 11 99 45 9 Language 5420 0 Language lo Ota Lae NO LI D OE 30r ak lg 2 64 96 42 48 28 76 Sensory 5435 2 Sensory de 696 50 tod 7 38 00 MOTQI O4 0 3 MOUOES 4 HA By DO 22 95 92 12 Ot er 5159 A Other 11 1 3 Save file as Pxxx CSM foci in PXxx SURFACES directory 11 2 Create coordinate color file 11 2 1 Use text editor to create color file RGB with following format BeginHeader comment date Fri Feb 04 2005 encoding ASCII EndHeader Language 0 255 0 Area 01 green Motor 255 0 0 Area 02 red Sensory 0 0 255 Area 03 blue Other 25
32. volume to assure the entire VOI is visualized Readjust the slider bars if needed and re select the Crop button to restore the desired sub volume Switch to the parasagittal panel and scroll to a slice where the partially cropped image volume is maximal in extent Adjust the min and max Z slider bars to the desired limits Once satisfied with the defined VOI in all 3 planes select Save 76 2 Volume Preparation continued 2 5 9 Select the hemisphere and region for the newly cropped volume in the Enter Cropped Volume s Extent Settings window Typically select Hemisphere left and Region entire cerebral hemisphere one or both 2 5 10 Select Save with default file name in the Save Volume As window 2 5 11 Typically we will not deal with Identify Cut Faces 2 6 Select Set Peaks 2 6 1 Inthe Hist window use the left mouse button to move the red bar to the left most peak of the histogram If the histogram does not have a clear gray matter peak select a value that results in roughly half the gray matter voxels being above threshold appearing green in the volume window 2 6 2 Select Set Gray Matter Peak button 2 6 3 Inthe Hist window use the left mouse button to move the red bar to the right most peak of the histogram If the histogram does not have a clear white matter peak select a value that results in roughly half of the white matter voxels being above threshold appearing green in the volume window 2 6 4 Select Set White
33. within and outside the labs in which the tool was created Other tools are discussed in Section 1 4 In Section 6 we discuss possible future work of evaluating other methods to provide further insight into how each method impacts results as well as the expected accuracy efficiency and distinct benefits of each method We selected a method from each of the two categories discussed in Section 1 as representative samples of each approach 2 4 1 Surface based anatomical normalization methods Caret http brainmap wustl edu caret Caret is a software tool developed by David Van Essen Heather Drury and John Harwell at Washington University Options for surface based transformation allow for the source to be deformed to the target while constrained by explicitly designated landmarks called Core6 landmarks Core6 includes the fundi of the calcarine sulcus central sulcus and lateral fissure the anterior half of the superior temporal gyrus STG and the medial wall cortical margin split into dorsal and ventral portions These landmarks were selected on the basis of their consistency in location and extent Caret deforms flat maps or spherical maps The spherical registration is more accurate and uses an algorithm developed by Bakirciogli et al Van Essen et al 2001 The basic strategy 1s to draw landmarks as prescribed by the Core6 guidelines on the source map then the landmark contours are resampled to establish corresponding numbers of landmark
34. 0 50 ylab mm col c grey50 orange blue FEE FE FE AE aE E aE AE aE E FE aE E aE a a E TE aE TE FE aaa aaa aaa LEE E E E H DELTA BETWEEN PRENORM AND POSTNORM DISTANCES HETE RE at E dd FE TE TE ae E E AE dg E TE FE FE FE TE TE AE FE E d TE ETE EEG normdelta caret lt expect postnorm avgDAB caret avgDAB normdelta spm lt expect postnorm avgDAB spm avgDAB delta results lt matrix c normdelta caret normdelta spm ncol 1 rownames delta results lt c CARET SPM2 barplot t delta results beside T main 3D Spread Reduction ylim c 0 5 ylab mm col c orange blue TETE FE FE Ae a dg dd AE FE TE aE A FE EEE EE aE CREATING DATA FRAME FOR STATISTICAL ANALYSIS TETE ae E ae dg dd AE TEE aE EEE AE E E E E EE E E E E E E make id pairs function brainIDs index lt outer brainIDs brainI IDs lt JTdl amp ouvcer DrarnrDs brarnriDs function d Ii T Id2z ouvcer bbralmnrDs bDrarniDs runcc5omi rtm 1 data frame idl id1 index id2 id2 index jackknife estimate of variance ids lt make id pairs brainlistSBrain ID uniqueids lt unique brainlistSBrain ID differences DAB vec lt caret DAB vec spm DAB vec jackknife diffs sapply uniqueids function i mean differences DAB vec idsSidl i K ids 1g2 e3 mean differences mean differences DAB vec se differences lt sqrt var jackknife diffs 10 ci mean lt mean differencestc 1 96 1 96 se differences
35. 1 language sites in 11 subjects 20 SPM2 P54 PSTG 0 5 2 sulcus P54 35 PSTG P55 4 MSTG no MMTG no MMTG P60 29 MSTG 1 3 no MMTG 1 3 no MMTG MSTG P61 25 0 5 2 sulcus 1 3 no MMTG P61 29 PMTG l l P62 33 OpIFG l l P63 25 MSTG 0 5 2 sulcus l P117 33 PSMG l sulcus Plod l 40 ASMG P1176 28 MSTG l l In order to refine the spread reduction analysis we removed all incorrectly mapped language sites resulting in a data set of 11 sites We then ran the spread reduction calculation again with the revised data set see table 14 The results are summarized in figures 28 and 29 and table 13 56 2D Language Site Spread 2D Spread Reduction lt P a em D w E e E F E E gis S Expected CARET SPM2 CARET SPM2 Figure 28 2D analysis of mean distance between the 11 correctly mapped language sites 3D Language Site Spread 3D Spread Heduction m ux S T m e E E E S e i L e D e Expected Caret SPM 2 CARET SPM2 Figure 29 3D analysis of mean distance between the 11 correctly mapped language sites 57 Table 13 Refined spread reduction results with the 11 correctly mapped language sites 2D Spread Reduction 3D Spread Reduction n 11 Caret 9 9 mm 3 2 mm SPM2 4 8 mm 1 3mm Table 14 Summary of 11 correctly mapped language sites ID Fe3 t Location The refined spread reduction analysis revealed an improvement for the Caret normalization
36. 24 2 65 59 SPM2 54 1 18 65 57 29 98 19 77 SPM2 54 1 19 44 42 25 27 49 78 SPM2 54 4 2 53 19 11 6 45 47 SPM2 54 1 20 63 22 56 81 28 92 SPM2 54 5 21 65 78 SAT 21 85 SPM2 54 4 22 59 66 2 94 14 9 SPM2 54 4 23 60 49 7 31 8 19 SPM2 54 2 24 59 46 0 5 8 86 SPM2 54 4 25 47 49 15 57 52 16 SPM2 54 4 26 55 67 13 5 41 02 SPM2 54 2 3 46 96 1 48 55 01 SPM2 ES Appendix G Deformation Map File BeginHeader comment Deformed with CARET v5 11 date Sun Jun 6 15 26 44 2004 encoding ASCII EndHeader deform map file version 2 flat or sphere DEFORM SPHERE deformed file name prefix caret source directory source spec Human colin L REGISTER with POP ATLAS xxxxx spec source landmark border Human colin L LANDMARKS REG with INDIVIDUAL COREO xxxxx borderproj source closed topo Human colin Cerebral L CLOSED xxxxx topo source cut topo Human colin Cerebral L CUTS xxxxx topo source fiducial coord Human colin Cerebral L FIDUCIAL TLRC 711 2B xxxxx coord source sphere coord Human colin Cerebral L SPHERE STD xxxxx coord source deform sphere coord deformed 4K NoFidHuman colin Cerebral L SPHERE STD xxxxx coord source deform flat coord source flat coord Human colin Cerebral L FLAT CartSTD xxxxx coord source resampled flat coord source resampled deformed flat coord source resampled cut topo sphere resolution 4610 border resampling 2 8 000000 spherical number of cycles 3 smoothing parameters 0 1 000000 100 20 10 30 morphing parameters 0 1
37. 2941 117 7 1 3 3 This information is needed to calculate the nelements argument for the mincresample function Calculate the nelements argument as follows dims out int round dim pixdim in 1 0 dim 256 pixdim_in 0 892941 nelements lt int round 256 0 892941 1 0 229 This calculation has been built into the Excel spreadsheet Resample PatientBrainData xls seen in Appendix B 1 3 4 Call the resample command and create a volume in 1 mm cubic voxels mincresample clobber nelements 229 229 229 step 1 0 1 0 1 0 Pxxx Exxxxx Sx mnc Pxxx Exxxxx Sx lll mnc 73 1 Patient MRI orientation and preparation continued 1 4 If functional MRI fMRI images are available for a subject it may be resampled as well and included in the normalization process 1 5 Open SureFit in PX XX directory on the command line 1 6 Read in the 111 mnc volume by selecting Volume Operations Read Volume 1 from the menu bar and selecting the desired mnc file If fMRI file is used load that file in as Vol 2 If the MR image is not centered in the Vol 1 window left click on the image and drag to the desired location 74 2 Volume preparation 2 1 To prepare the volume data for the segmentation process select SureFit Volume Preparation from the menu bar There are 7 tabs in the Volume Preparation window Typically we will use the first five tabs as follows 2 2 Select Volume Information 2 25 2 2 2 212 3 2 2 4 2 215 2 2 6 2 25 f
38. 3 p 10 For subject p the average distance between language sites in subject p and subject p prior to B normalization p E EE L leal lt je PO pi a d The average distance between language sites in subject p and subject p post normalization R pi pp nv pj where B individual left hemisphere n number of language sites identified for subject p m number of language sites identified for subject p D Euclidian distance between points X is a variable representing site coordinates prior to normalization Y is a variable representing site coordinates after normalization Euclidian distance between points was measured according to d 2 this equation D a b i 1 ay bk where d 2 or 3 depending on whether the coordinates are 2D or 3D 4 Calculating Distance Across Subject Pairs Within the language functional region Let B left hemisphere p patient 1 orange nz3 p patient 2 purple me 2 Y variable representing site coordinates D Euclidean distance Figure 15 Calculating average distance between language sites across two subjects represented by the orange and purple sites respectively Figure 15 illustrates the distance calculation using three language sites of one subject orange and two language sites from another subject purple Our distance measure used only the distances between the orange and the purple sites We believe that the combina
39. 371 Z Algorithm 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 Caret 0 SPM2 0 SPM2 0 SPM2 0 SPM2 0 SPM2 0 SPM2 0 SPM2 0 SPM2 0 SPM2 0 SPM2 0 SPM2 0 SPM2 0 SPM2 0 SPM2 0 SPM2 0 SPM2 121 122 Excerpt from PostNorm csv Brain CSM Site ID Region Number X Y Z Algorithm 54 3 10 63 82 7 532 12 588 Caret 54 3 12 56 92 5 503 41 632 Caret 54 3 13 50 01 17 85 33 39 Caret 54 3 14 49 45 19 31 44 795 Caret 54 3 ibe 42 86 17 11 59 118 Caret 54 3 16 52 8 30 78 57 216 Caret 54 4 18 49 3 36 85 8 4897 Caret 54 1 19 38 97 26 66 43 719 Caret 54 4 2 40 43 14 97 36 867 Caret 54 1 20 52 87 52 62 20 228 Caret 54 5 21 47 73 40 58 10 555 Caret 54 4 22 60 39 9 072 9 5786 Caret 54 4 23 56 46 12 82 8 5118 Caret 54 2 24 57 29 10 37 3 5565 Caret 54 4 25 46 61 14 55 48 95 Caret 54 4 26 39 67 16 87 36 897 Caret 54 2 3 47 06 3 755 47 295 Caret 54 l 30 56 81 55 13 16 654 Caret 54 5 31 56 41 50 59 21 538 Caret 54 l 35 62 86 53 79 16 159 Caret 54 1 36 47 97 60 92 27 693 Caret 54 4 37 57 95 11 69 19 075 Caret 54 5 42 63 43 53 08 11 23 Caret 54 5 43 52 09 24 79 1 809 Caret 54 3 10 65 18 12 26 17 62 SPM2 54 3 12 56 62 5 31 44 33 SPM2 54 3 13 59 22 16 7 47 51 SPM2 54 3 14 51 74 15 27 55 97 SPM2 54 3 15 41 48 16 66 63 53 SPM2 54 3 16 44 69
40. 4 Following normalization using both methods the neuroanatomist expert viewed cortical flat maps including sulcal depth patterns and cortical sites via the Caret GUI The SPM2 normalization was viewed on the left side of the screen and the Caret data on the right side of the screen Each site number was identified via a mouse click The neuroanatomist identified the normalized location of each mapped site based on CPS The author recorded the post normalization parcellations and compared them to the pre normalization parcellations assigning a score and or error type A correct mapping received a score of 1 Error types and scores were assigned as follows O Figure 18 Example of Type I Error Site is mis mapped to a parcel across an subjective boundary as Type 1 Error The site 1s located in an incorrect parcel across a subjective boundary and receives a score of 0 25 Figure 18 illustrates this error type Type 2 Error The site 1s located in the sulcus adjacent to the correct parcel and receives a score of 0 5 Figure 19 illustrates this error type Type 3 Error The site 1s located in the incorrect parcel across a sulcal boundary and receives a score of 1 Figure 20 illustrates this error type b dd Type Error Q9 iud AST E PoISTG AMTG A a Peiter delineated by a dashed line Figure 19 Example of Type 2 Error P117 Pre norm sulcal depth flat map with cortical sites on the left
41. 4 329 346 Kochunov PV Lancaster JL Fox PT Accurate high speed spatial normalization using octree method NeuroImage 1999 10 724 737 Lancaster JL Summerln JL Rainey L Freitas CS Fox PT The Talairach Daemon a database server for Talairach Atlas Labels Neuroimage 1997 5 4 S633 Lee T C Kashyap R Chu C N Building skeleton models via 3 D medial surface axis thinning algorithms Graph Models Image Processing 1994 56 462 78 Le Goualher G Prock E Collins L et al Automated extraction and variability analysis of sulcal neuroanatomy IEEE Trans Med Imag 1999 18 206 216 Liu T Shen D Davatzikos C Deformable registration of cortical structures via hybrid volumetric and surface warping NeuroImage 2004 22 1790 1801 Lorenson W Cine H Marching cubes a high resolution 3 D surface reconstruction algorithm Comput Graph 1987 21 163 9 Mangin JF Riviere D Cachia A et al Object Based Strategy for Morphometry of the Cerebral Cortex IPMI Ambleside UK 2003 May A Ashburner J Buchel C et al Correlation between structural and functional changes in brain in an idiopathic headache syndrome Nat Med 1999 5 7 836 838 70 Minoshima S Koeppe RA Frey KA Kuhl DE Anatomic standardization linear scaling and nonlinear warping of functional brain images J Nucl Med 1994 35 1528 1537 Modayur B Jakobovits R Maravilla K Ojemann G Brinkley J Evaluation of a visualization based approach to functional bra
42. 5 255 255 Area 04 white Stim 0 0 0 Area 05 black 11 2 2 Save as CSM focicolor in Pxxx SURFACES directory 11 3 View magnet space coordinates on individual surface 11 3 1 Open Caret in Pxxx SURFACES directory 11 3 2 Select Pxxx Exxxxx Sx_111 L REG with Colin_Core6 xxxxx spec Make sure there are no foci focicolor files selected select the inflated surface and accept the remaining default selections The fiducial should be the magctr version 11 3 3 Open Data File Foci Color File COLIN L LANDMARKS REG with INDIVIDUAL COREO CSM focicolor 99 11 Create CSM coordinate files continued 11 3 4 11 3 5 11 3 6 11 3 7 11 3 8 11 3 9 Open Data File Foci File Pxxx CSM foci Press D C and toggle on the Foci checkbox at the bottom Press L for a lateral view of the CSM sites The CSM sites should appear centered around the root of the sylvian fissure Select Layers Foci Project Fiducial Foci Hemisphere only keep offset from surface setting File Save Data Files Foci Projection Files Pxxx CSM fociproj Switch to the flat map view 11 3 10 Save flat coordinate file as follows 11 3 10 1 File Save Data File Foci File 11 3 10 2 Foci Associated with Surface Type Flat 11 3 10 3 Filename Pxxx CSM flat foci 11 3 11 Select D C Surface Shape Depth to switch from viewing folding patterns to sulcal depth patterns 11 3 12 Select D C Foci menu Draw Foci as Spheres adjust foci size as desired 11 3 13 Sel
43. 5 l OpIFG P63 MITG 0 25 l MMTG P63 ASTG 1 3 no MMTG P164 MSTG 1 3 no MMTG VPRG 1 3 no VPoG P164 P176 25 OpIFG 0 25 l T FO Seventeen errors were unique to the SPM2 spatial normalization method representing over 27 of all SPM2 errors Error analysis reveals that the most common region to be erroneously mapped was the ventral part of the precentral gyrus VPrG as more than 29 of unique errors should have been mapped to this parcel There are four other parcels that each account for more than 10 of SPM2 unique errors middle part of the inferior temporal gyrus MITG anterior part of the superior temporal gyrus ASTG middle part of the superior temporal gyrus MSTG and opercular part of the inferior frontal gyrus OpIFG Type 3 errors accounted for 59 of the unique errors with the remaining errors being type 1 There were no type 2 errors unique to the SPM2 method 55 Language Site Localization Localization accuracy for language sites mapped by Caret was 72 62 versus SPM2 accuracy of 84 52 Caret incorrectly mapped 9 42 9 of the 21 language sites SPM2 incorrectly mapped 6 28 6 Again we observed that the superior temporal gyrus STG was the most problematic for both methods with seven of the ten incorrectly mapped language sites being located on the superior temporal gyrus STG Five language sites were incorrectly mapped by both methods Table 12 Summary of mapping by both methods of 2
44. 9 mL ele Save as Pxxx CSM stripped 109 Appendix D norm_coord m script LUNCCTON norm coord tintsle P spmgebtl 1y BL yoo Lect dero tmqtron Pal repna eP go yey is Va te Be 2 6 is a V spm vol P tire Spm koad4 anita 3 out inname SCrCat opm e SE LGT outfile fopen outfname w fprintf Writing output file ss dXn outfname length tile T for i l length file xetile r 2 43 y file 1 3 z file i 4 e ps ps vx inv V l mat c 1 ZS The voxel in the deformation to sample normx spm sample vol V 1 vx 1 vx 2 vx 3 1 normy spm sample vol V 2 vx 1 vx 2 vx 3 1 normz spm sample vol V 3 vx 1 vx 2 vx 3 1 forinti edini lji ftorqvmtrtoutrfile Sd S2 S42 cede c eL norms mnormvnormz end fclose outfile 110 Appendix E merge spm foci sh script bin sh if ne 2 then echo usages Merge spm lOc lsh Spm TOCIL coords Exe magrocl exit 1 EL SPMFOCI 1 MAGFOCI 2 HEADER 1 cat SMAGFOCI while read line do if SHEADER eq 1 then echo Sline else FOCINUM echo Sline cut f1 d COORD grep SEKOCINUM SSPMEOCI eut E2 4 g REST echo Silane cut f5 6 d echo SFOCINUM COORD SREST fi hit echo Sline grep BEGIN DATA If p shave T s then HEADER SO w aka done 111 Appendix F Spread reduction source code and input files HH HH Ht HH HH HEE HH HEE HH HEE HH HEE HH HE HH HE
45. A Cortical surface based analysis II inflation flattening and surface based coordinate system NeuroImage 1999 9 195 207 Fox PT Mintun MA Reiman EM Raichle ME Enhanced detection of focal brain responses using intersubject averaging and change distribution analysis of subtracted PET images J Cereb Blod Flow Metab 1988 Oct 8 5 642 653 69 Gee JC Reivich M Bajcsy R Elastically deforming 3D atlas to match anatomical brain images J Comput Assist Tomogr 1993 Mar Apr 17 2 225 236 Gee JC Barillot C Le Briquer L Haynor DR Matching structural images of the human brain using statistical and geometrical image features Proc SPIE Visualization in Biomedical Computing 1995 Aug 3573 278 289 Hill DLG Batchelor PG Holden M Hawks DJ Medical Image Registration Phys Med Biol 2001 46 R1 R45 Joshi SC Miller MI Christensen GE et al Hierarchical brain mapping via a generalized Dirichlet solution for mapping brain manifolds Proc SPIE Vision Geometry IV 1995 Aug 2573 278 289 Joshi S Brain Mapping lt http cis jhu edu sarang brain html gt Accessed Dec 2003 Kikinis R Jolesz FA Lorensen WE et al 3D reconstruction of skull base tumors from MRI data for neurosurgical planning Proceedings of the Society of Magnetic Resonance in Medicine Conference 1991 Kriegeskorte N Goebel R An efficient algorithm for topologically correct segmentation of the ortical sheet in anatomical MR volumes NeurolImage 2001 1
46. BY BRAIN ID HETE aE dg dd Hea ae ae E TOTO dd dg dd dd dd FE TE ERE ETT T lang prenorm list split langcoord prenorm langcoord prenorm Brain ID vd dg dd dg dd dd dg dd dg OEC d dd dg dd d dg T ATTEMPT TO GET BRAIN LANG SITES IN CORRECT ORDER aE aE dg dd Ae a ae ae A AE a AEE AEE aE Ea EAE d dd dg EE aaa aaa braunlise lt Langcooro prenormlc Braan LD XW May A length lt dim brarnl Istr t brainID 0 row U tempNumRow U brainlistlengths vector length 0 for m in 1 length if row 0 brainID lt brainlist m Brain ID row 1 else if brainlist m Brain ID brainID row row 1 else brainlistlengths lt c brainlistlengths row row 1 brainlD lt brainlist m Brain ID END OF FOR LOOP brainlistlengths is a vector that lists the number of lang sites per brain in the order the brains will be in calcdist xx matrices brainlistlengths lt c brainlistlengths row 118 HEE EE HHH HH EEE EEE HE HH EE EE EHH HE EE EE OH EE HE EEE OH HE EE EEE EE EE EE EE HEE EE Er Ep Ep Ep EE CALCULATION OF NUMBER OF LANGUAGE SITES FOR EACH BRAIN AND TOTAL NUMBER OF ROWS TO BE USED LATER HEE EH HH HEH EEE EEE HE HE EE EEE EEE HE EE OH EHH FE EEE HE HE EEE OH OH EE FE EEE EE EE EEE HE E length length lang prenorm list totalLength 0 RowNum vec lt vector length 0 FOr Gk ay Lsrtength 4 tempRowNum lt nrow lang prenorm list k tempLe
47. Border Delete border with mouse as needed to delete a bad border and redraw Layers Border Project border nearest tile 8 20 10 File Save Data File border projection file PXXX EXXXXX SX_111 L LANDMARKS forReg with Colin Core6 xxxxx borderproj 94 8 Drawings borders in Caret for surface normalization continued 8 20 11 Switch to the SPHERICAL surface 8 20 12 D C Surface Miscellaneous Hide Surface 8 20 13 Toolbar View and rotate the invisible surface making sure the landmarks look smooth and curvy with no sharp turns or hooks e g if a border point somehow got translated to the origin 8 20 14 Toolbar Spec and select Border from the spec selection menu Click X next to the PXXX EXXXXX SX 111 L LANDMARKS FromFlattening xxxxx borderproj entry to remove this entry from the spec file Make sure PXXX EXXXXX SX III L LANDMARKS forReg with Colin Core6 xxxxx borderproj is the only borderproj entry and there are no border entries keep ForSPHERICAL REGISTRATION Human Class3 bordercolor as a bordercolor entry 95 9 Caret normalization of individual surface to atlas 9 1 Now we are ready to run the normalization algorithm Select Surface Deformation Run Spherical Surface Deformation 9 1 1 Individual tab 9 1 1 1 Spec PXXX EXXXXX SX III L REG with Colin Core6 xxxxx spec 9 1 1 2 Border PXXX EXXXXX SX III L LANDMARKS forReg with Colin Core6 xxxxx borderproj 9 1 1 3 Closed Topo PXXX EXXXXX SX 111 L full seg
48. Caret a spherical registration algorithm used landmark borders to create a deformation map SPM2 spatially normalized the individual volume image to the avg152T1 Minc file to create a deformation file which was aligned to ICBM152 space The deformation for each method was applied to the individual coordinate file in magnet space 35 coordinates resulting in a normalized coordinate file The result was a set of coordinate files registered to the same reference space ICBM152 space These normalized coordinate files were used to evaluate accuracy of each method based on spread reduction between sites and preservation of anatomical localization Step 6 Of the 198 CSM sites we were especially interested in the 21 sites identified as statistically significant for naming errors see table 4 Such sites have been found within and outside areas classically considered language function regions We believe that a hidden pattern of language production exists that could be revealed with the help of spatial normalization Statistical significance was derived by analysis of the patients responses Analysis included comparing the patient s pre surgery test responses to the intraoperative test responses To determine whether naming disruption at a site determined by the neurosurgeon was an effect of stimulation or attributable to the baseline naming error rate of the subject a within subject analysis of naming errors was carried out Fischer s exact t
49. DMARK SF STSant from the Name menu and click Apply 8 19 7 Starting at the dorsal tip of the superior temporal gyrus closer to the center of the surface trace along the white line until you reach the end point then shift click 8 19 8 If you re not happy with a border select Layers Border Delete border with mouse as needed to delete a bad border and redraw 8 19 9 Layers Border Project border nearest tile 8 19 10 File Save Data File border projection file PXXX EXXXXX SX _ 111 L LANDMARKS forReg with Colin Core6 xxxxx borderproj 92 8 Drawings borders in Caret for surface normalization continued 8 19 11 Switch to the SPHERICAL surface 8 19 12 D C Surface Miscellaneous Hide Surface 8 19 13 Toolbar View and rotate the invisible surface making sure the landmarks look smooth and curvy with no sharp turns or hooks e g if a border point somehow got translated to the origin 8 19 14 Toolbar Spec and select Border from the spec selection menu Click X next to the PXXX EXXXXX SX 111 L LANDMARKS FromFlattening xxxxx borderproj entry to remove this entry from the spec file Make sure PXXX EXXXXX SX III L LANDMARKS forReg with Colin Core6 xxxxx borderpro is the only borderproj entry and there are no border entries keep ForSPHERICAL REGISTRATION Human Class3 bordercolor as a bordercolor entry 8 19 15 Starting at the ventral tip of the central sulcus near the origin scale bar near the center of the sur
50. E Lang scarey list x split Vengecord caret langcoord caret tr ID postnorm caret list x lapply lang caret list function cd dieL caretimatrlx lt cbindicdSx cdoY casZ dist dist caret matrix method euclidean Si ttt tt HE EH HE HH FE EH HEE EH EE EH HH EEE EH HE EE EE EE EE EH HE EE SPM2 DISTANCES BETWEEN LANGUAGE SITES W in EACH BRAIN ttt tHE EH HE HEH FE EH HEE EH EE EH EEE EEE HE OE EE EE EE EH HH EE latg spm list lt spliti landcoord spm langcoord spmsBrain rb positnormispm lrst Lapoly Lang sSpm LIST functriondicd 4 disi spm matrix lt cbanog cds GdoY c092 dist dist spm matrix method euclidean e tt Ht HH HHH EH HE HE HH HH HEE EH EE EH HE TE EE HE HE AE TE EE HE HH EEE EE EE E E E dd DISTANCE ACROSS BRAINS FUNCTION DAB tt Ht tH HH EH HE HE HH HH EEE EE TE EE HE EEE EH HE EEE E EE EE HHH EE HE DAB lt function brainlDs distmatrix index lt outer brainIDs brainIDs lt DAB vec lt distmatrix index DAB vec lt as vector DAB vec DAB vec 114 HEE EHH HH HH EE EEE HH HH EE EH HE HE EE EH OH EE EEE EE OE EE EE EE HE EE EE AVG PRENORM DISTANCES FOR LANGUAGE SITES ACROSS ALL BRAINS HEHE HHH HH EE EHH EH HH EE EE EE HE EE EH HE EE FE EE EH EE EE EEE EE EE EE PRENORM DATA SET OF LANGUAGE SITE COORD FOR ALL BRAINS COOPQO prenorm mebrix lt Langoourd prenorm oc Ux Ay MZ PRENORM DATA SET OF DISTANCES BETWEEN LANG SITES IN w in and ACROSS BRAI
51. E HH HEE HH HEE HH HEE HH HE HH EE CH HE CH HH EE HH HE CALCULATING AVERAGE 2D EUCLIDIAN DISTANCE BETWEEN LANGUAGE SITES ACROSS 11 HUMAN BRAINS DATE 2 25 2005 AUTHOR VERONICA SMITH HH HH Ht HH Ht EE HH EE HH HE HH HEE HH EE HH HE HH HEE HH HEE HH HEE HH HE HH EE HH HE HH HE CH HH EE A HERE HERE HERE E EE EHH E E E PREP WORK EHE EHE HERE E EE EHH EH HEE CLEAN UP rm list l1s all TRUE IMPORTING DATA pre dat lt read csv FlatPreNorm csv postdat lt read csv FlatPostNorm csv pp dar lt x rbind pre dar y post dat SUMMARIZING DATA summary pre dat summary post dat summary pp dat LOAD 3D PLOTTING PACKAGE library scatterplot3d library help scatterplot3d itt HH HH HH Ho HH HOH HH OEE OE EE T PRENORM LANGUAGE SITE COORD itt HH HH HH HHH HH d dt HO EE HOE EE T langcoord prenorm pre dat 0 temp matrix langcoord prenorm length dim pre dat 1 for i in 1 1length if pre dat i CSM Region 1 temprow pre dat i temp matrix lt rbind temprow langcoord prenorm langcoord prenorm temp matrix 4 END OF FOR LOOP 4 HEHEHE EE HEH EEE HEH HEE HH HEE HH HE EF CARET LANGUAGE SITE COORD HEE HEH HEE HH HEE HHH EE HH HEE HH HE EF Langeoord carer lt post datl0 temp matrix lt langcoord caret length lt dam post dat 1 for a cm Iilength d ie C pOSt dauli CSM Region 1 amp amp post ocdatli Albgorichim
52. IN BOTH BRAINS caLlcdist spm matrix lt as matrix caLbcoeirst spm spm b1 DS lt La a gecoord CaretSBrain ID spm DAB vec lt DAB spm bIDs calcdist spm matrix spm avgDAB lt mean spm DAB vec HHH HEH EH HEH EH HE HEH EH EH EH RE ERE RE E RE ERE RE ERE ERE RE E E ER E B E PRESENTING RESULTS IEEE EEE HE HEH EE HH EE ERE ERE HOH EH EH EH RE ERE RE E EH EH EE E B E par mfrow c 1 2 prelim results lt matrix c expect postnorm avgDAB caret avgDAB spm avgDAB ncol 1 rownames prelim results c Expected CARET SPM2 barplot t prelim results beside T main 2D Language Site Spread ylim c 0 100 ylab mm col c grey50 orange blue 115 TOTO HE aE AE aE TOO TU EO TEE TEE aaa E E E H DELTA BETWEEN PRENORM AND POSTNORM DISTANCES HETE aE dg dd Ae a ae ae a E a aE aE AEE aE EEE d dudd dg dd dg dr normdelta caret lt expect postnorm avgDAB caret avgDAB normdelta spm lt expect postnorm avgDAB spm avgDAB delta results lt matrix c normdelta caret normdelta spm ncol 1 rownemes delta results lt c CARET SPM2 barplot t delta results beside T main 2D Spread Reduction ylim c 0 10 ylab mm col c orange blue Hake rd perrse tunotrzon brsintDs lt 4 index lt outer brainIDs brsinlbs lt ELE braint Ds Drainrbs icunoctconm rx Td2s outeri brarmiDs bDrauinlblDs runcctromirJJ 3 data frame idl idl index 1d2 1id2 inde
53. Matter Peak button 2 7 We will not typically adjust the parameters using the List Parameters tab Using the Resampling tab is not recommended for normal segmentation 2 8 When complete with all volume preparation select Save and Close E 3 Segmentation surface generation and automated error correction 3 1 Make sure that the structural MRI volume to be segmented is loaded as Vol 1 If there is a functional MRI it should be loaded as Vol 2 3 2 An initial segmentation is run first to determine if bias correction or other pre processing is needed Select SureFit Run SureFit A notebook window appears with 3 tabs Run SureFit Interactive Error Correction and A La Cart In the Run SureFit tab use the following selections e Segmentation Scope Extract Cerebrum Segment e Fill Ventricles Yes Leave Keep intermediate files unselected 3 3 Check for segmentation quality once initial segmentation is completed as follows 3 3 Select the Interactive Error Correction tab 3 32 Inthe window press the Update Handle Count button to determine the number of topological errors handles for the volume loaded in Vol 2 Note Occassionally this method can yield a few 1 3 false positives small handles in the volume that do not appear in the surface reconstruction If Handle count number returned is greater than or equal to 15 then it is a good idea to consider some additional pre processing to upgrade image quality before moving forward w
54. NS calcdist prenorm lt dist coord prenorm matrix method euclidean calodrstopcenorfmumdgtr x lt as matrix CalCdist porenorm prenorm bIDs langcoord prenorm Brain ID prenorm DAB vec DAB prenorm bIDs calcdist prenorm matrix prenorm avgDAB lt mean prenorm DAB vec expect postnorm avgDAB lt sqrt 1 3085 prenorm avgDAB TOTO EE TOTO OE E FE E TO FE AE FE FE TOO TOTO TOO TEE ETT AVG CARET DISTANCES FOR LANGUAGE SITES ACROSS ALL BRAINS TETTE ae A Ae a TOTO TTE OEC TOTO ETE TOTO dd d dg dd dg d dd ETE coord caret matrix langcoord caret c X Y Z E POSTNORM CARET DATA SET OF LANG SITE COORD FOR BOTH BRAINS 4 calcdist caret lt dist coord caret matrix method euclidean 4 POSTNORM CARET DATA SET OF DISTANCES BETWEEN LANG SITES IN BOTH BRAINS calcdist caret matrix as matrix calcdist caret caret DIDS lt langcoord carecsoBrain rD caret DAB vec lt DAB caret bIDs calcdist caret matrix caret avgDAB mean caret DAB vec TETTE FE FE E TE FE AE E TETTE OE EEE TE TE FE AE TE FE AE AE E TE TE RE E E E E E E EE E EH AVG POSTNORM SPM2 DISTANCES ACROSS ALL BRAINS HEFE aE at E dd E TE TE AE AE TETTE TE TE E dg AE E TE TE TE E TE FE TE E E EET T coord spm matrix langcoord spm c X Y Z POSTNORM SPM DATA SET OF LANGUAGE SITE COORD FOR BOTH BRAINS T calcdist spm dist coord spm matrix method euclidean 4 POSTNORM SPM2 DATA SET OF DISTANCES BETWEEN LANG SITES
55. This template is used as a standard template in the MNI brainweb simulator ICBM Probabilistic Atlases Arthur Toga Laboratory of Neuro Imaging LONI Director and John Mazziotta UCLA Brain Mapping Center Director and principal investigator of ICBM lead a team of researchers who have created a variety of probabilistic atlases as they work to achieve the team s ultimate goal of a four dimensional atlas and reference system that includes both macroscopic and microscopic information on structure and function of the human brain in 7 000 persons between the ages of 18 and 90 years As discussed the fact that no single unique physical representation for the human brain is representative of the entire species the variance must be encapsulated in an appropriate framework Mazziotta and Toga have chosen a probabilistic framework in which intersubject variability is captured as a multidimensional distribution Accessing data from a probabilistic atlas will produce a probability estimate of structures and function based on the distribution of samples obtained This frameworks also differ from frameworks like the ICBM152 which is an average brain space The average brain framework is created using a density based approach An atlas using the density based approach is an average space constructed from the average position orientation scale and shear from all the individual subjects It is therefore both an average of intensities and of spatial posit
56. VA colin27 volume area ASA average surface area of 11 subjects AVA average volume area of 11 subjects CVA AVA 1 4407 CSA ASA 1 3085 In order to accurately estimate EPoD we considered that linear distances do not increase linearly with the increase of volume or surface areas A linear dimension will increase by the square root of M as the surface increases by M times In the case of a volume a linear dimension will increase by the cubed root of N as the volume increases by N times We used the following equations to calculate EPoD in 2D and 3D space 2D space EPoD NCSA ASA x Pr D 3D space EPoD Y CVA AVA x PrD 43 Having normalized for the increase in surface and volume area of the target atlas we were prepared to compare the actual post normalization distances to the expected values We calculated average distance between language sites in both 2D space and 3D space Thus we used two different coordinate systems For 2D space we used the Caret coordinate space which sets the ventral tip of the central sulcus as the origin white cross in Figure 16 and Figure 17 on the flat maps and for 3D space we used the ICBM152 coordinate space Figure 16 P54 language sites mapped in 3D left and 2D right space Figure 17 P117 with language sites mapped in 3D and 2D space 44 Anatomical Localization To measure how well anatomical normalization methods preserve anatomical location we used CPS as outlined in step
57. Voyager software was developed by Rainier Goebel Maastricht University originally introduced as a tool for analysis and visualization of functional and structural imaging data in 1998 It is now a commercial software package featuring cortex based inter subject normalization based on gyral sulcal patterns of individual brains as well as other functions listed previously Goebel 2000 2 4 2 Volume based anatomical normalization methods Analysis of Functional NeuroImages AFNI http afni nimh nih gov afni about descripadfaad AFNI is a software environment for processing and displaying functional MRI data on an anatomical substrate It was designed and written at the Medical College of Wisconsin primarily by Robert Cox now director of scientific and statistical computing core at the National Institute of Mental Health It is a free software package that uses a base unit of data storage called the 3D dataset which consists of one or more 3D arrays of voxel values with some control information stored in a header file AFNI s spatial normalization feature requires the user to select various markers first to align the anterior commisure and posterior commisure and a second set of markers to define the bounding box of the subject s brain Then a 12 sub volume piecewise linear transformation to Talairach coordinates is performed for both anatomical and functional volumes Cox 1996 SPM http www fil ion bpmf ac uk spm Karl Frist
58. We also support this type of functionality in spatial normalization tools like Caret and SPM2 We demonstrated that the surface based method allows for more quantitative and qualitative assessment of tool performance than does the volume based method This evaluation led to a deeper understanding of the limitations and advantages of each method and provides a frame work that can be used with modification to determine spatial normalization accuracy 64 Language Localization Patterns Figure 34 illustrates the cortical parcellation of 16 language sites as mapped by Caret and SPM2 The five sites that were incorrectly mapped by both methods are excluded from this illustration Caret Language Site Localization aem Pun Oh d Most common parcels for language sites highlighted in green Figure 34 Mappings of 16 language sites on the colin27 atlas with incorrect mappings circled in red 65 Section 6 Future Work Possible future work includes repeating the study with a larger sample using a probabilistic atlas as the target repeating the study using subject MRI data instead of CSM site data repeating the Caret normalization of the current data set using modified landmark guidelines incorporating standard metrics into the evaluation protocol and repeating the study to compare Freesurfer BrainVoyager AFNI a hybrid registration algorithm and a MMI registration algorithm As discussed in Section 2 3 we wou
59. al Institute MNI wanted to define a template brain that was more representative of the human population than the single brain used by Talairach and Tournoux They created a new template that was approximately matched to the Talairach brain via a two step process First they used 241 normal MRI scans and manually defined various landmarks and the edges of the brain Each brain was scaled to match the landmarks to equivalent positions on the Talairach atlas Second a sample set of 305 normal MRI scans from right handed male 239 and female 66 individuals were normalized to the average template of the first 241 brains using an automated 9 parameter affine transform From this MNI generated an average of the 305 brain scans This atlas is known as the MNI3065 atlas and was the first template created at MNI The current standard MNI template is named the ICBM152 because the International Consortium for Brain Mapping adopted this atlas as their standard template The ICBM152 atlas was created from an average of 152 normal MRI scans that were normalized to the MNI305 using a 9 parameter affine transform Brett 2003 23 colin27 Atlas A MNI lab member Colin Holmes underwent 27 MR brain scans These scans were then coregistered registered to each other and averaged to create a detailed MRI dataset of one brain The average of the 27 scans was then registered to the ICBM152 space to create what is called colin27 also known as the Colin atlas
60. ap representing the differences between the average sulcal depth of 10 epileptic and 12 normal subjects Dark areas represent where the epileptic subjects gyri are deeper and the white areas represent where the epileptic gyri are shallower than the 12 normal subjects 6l The colin27 atlas uncommon folding patterns of the supramarginal gyrus and lateral fissure are documented by Ono in the Atlas of Cerebral Sulci Four left terminal ascending segment patterns were delineated Upon inspection the colin27 folding pattern most closely matched the pattern illustrated in figure 15 9D in the Ono text It is described as follows a descending terminal portion which does not constitute the posterior transverse temporal sulcus This pattern occurred in 4 of the 25 autopsy specimen brains examined for variations and consistencies in location shape size dimensions and relationships to parenchymal structures Ono et al 1990 Two of the remaining three patterns represented 88 of the folding patterns found in this region with the final pattern representing 8 of the patterns found This gyral pattern impacts the sulcal pattern of the supramarginal gyrus SMG as SMG surrounds the posterior tip of the lateral fissure contributing to colin27 atlas unusual folding pattern in this parcel It is well known that anatomical variation between source and target can prove problematic for accurate registration This study supports previous findings that s
61. as widely used in international brain imaging studies and continues today to be the most widely used human brain 21 atlas Talairach space consists of 12 rectangular regions of the target brain that are piecewise affine transformed to corresponding regions in the template brain Using this transformation a point in the target brain can be expressed in Talairach space coordinates which allows for comparison to similarly transformed points from other brains Brinkley and Ross 2002 Today there is a database and data retrieval system named Talairach Daemon developed at the University of Texas San Antonio that performs the registration of target brains to the Talairach template brain http ric uthscsa edu projects talairachdaemon html This system returns anatomical labels using Brodmann area names for the cerebral cortex and other traditional feature based terms when queried with a stereotaxic coordinate from an individual subject s brain Thus the Talairach Atlas provides a symbolic representation textual of the brain The entire Talairach brain has been anatomically labeled using a five level volume based terminological hierarchy Level One hemisphere has six components left and right cerebrum left and right cerebellum left and right brainstem Level Two lobe divides each hemisphere into lobes or lobe equivalents In cerebrum and cerebellum lobes are as traditionally defined In brainstem three lobe equivalent
62. at maps volume averaged group functional data e g fMRI onto all 24 individual hemispheres in the atlas followed by spatial averaging across the individual maps yielding a population average surface representation that shows the most likely regions of activation and the maximal extent of plausible activation We selected the colin27 for primarily two reasons relating to the type of metrics we wished to use First we wanted to measure pre normalization and post normalization distances between language sites across brains both in 2D and 3D space If we were to use an average brain atlas like ICBM152 the blurring that occurs from averaging multiple brains would distort the flat map distances significantly after normalization because the sulci would become significantly shallower due to averaging as compared to the sulci in the individual flat maps Second evaluation of anatomical localization using CPS required visualization of the data on a single brain so as to determine if a site is indeed in the correct parcel The blurring of sulci and gyri that is a result of averaging individual MR images would make the evaluation very difficult 1f not impossible Given these constraints we selected the colin27 atlas that we received from the Laboratory of Neurological Imaging LONI at UCLA and registered it to MNI152 space using SPM2 25 2 4 Spatial normalization methods We considered five spatial normalization software tools that are commonly used
63. ate a formatted coordinate file with the following command merge spm foci sh spm Pxxx coord stripped Pxxx CSM foci gt spm Pxxx CSM foci Save this file in the Pxxx directory This foci file will serve as input into Caret for visualization of results 13 8 Move spm Pxxx CSM foci to the colin atlas directory e mv spm P117 foci J COLIN L LANDMARKS REG with INDIVIDUAL COREO 13 9 cd COLIN L LANDMARKS REG with INDIVIDUAL COREG caret5 13 10 Select caret P117 E10043 S4 111 L REG with Colin Core6 71785 spec and accept the default file selections with these exceptions toggle on INFLATED surface toggle on CSM focicolor make sure any foci or fociproj files are deselected select Human colin Cerebral L xxxxx surface shape not the deformed Pxxx surface shape file 13 11 File Open Data File Foci spm Pxxx CSM foci 13 12 Toolbar L to switch to lateral view 13 13 Toolbar D C and toggle on Foci 13 14 Layers Foci Project Fiducial Foci Hemisphere only keep offset from surface 13 15 File Save Data File Foci Projection spm Pxxx CSM focipro 106 13 SPM2 normalization of CSM coordinates to atlas continued 13 16 D C On the Shape drop down menu make sure Depth is the selected column if this option isn t available make sure you have Human colin Cerebral L 71785 surface shape loaded not the deformed P117 surface shape file 13 17 D C Foci menu Draw Foci as Spheres adjust foci size as desired 13 18 Switch to flat view
64. ation Methods essen 7 1 5 Cortical Stimulation Mapping and Visual Comparison Approach 12 IRE VDOC SIS ees runt du sea REM DENN M ME MM EE E MET ELM E 16 Secton 2b Survey EE te KEE 17 2 1 Surface Reconstr ecttOEi EE L7 2 52 Suridce Itt BIB EE d bot obtu tbid votis doit seco tui ect Adele 20 PAS A KT S RO oe 20 ZA Spatial OTS ST T EE 25 STEE R ee 28 BV CUNO ME E CSU S E A E Catus scone n dtt e ui etae oT EE NS E 46 DECOM DISCUSSION osos ou depo a oat teenie easels 58 BC CUO MNO T E saqads nau iM ente tau ET Md ME Ud RUE 65 LAS OP EE 67 Appendix A Evaluation Protocol oie trt dore dee 72 Appendix B Resampling MRL 423 eset ee e ened e te boe e teh geben 107 Appendix C Creating a Stripped Coordinate ke 108 Appendix R nor coord Ee EE 109 Appendix Emerge Spm eerste ees 110 Appendix E Spread Reduction Source EE 111 Appendix G Detormation Vp Tale catus dnb di dod ot e a Miet de e A ae 123 li LIST OF FIGURES Figure Number Page E STS LTR eT T 4 2 PUNCHONAL EENEG 6 3 rao perd iVe NOLO CLAD ceci Ev pa VM qam I E ME E IEEE eae 14 4 Skandia GUS Creens WOU E 15 a Cerebral Neocorte e ET EE 17 6 SureEit Cortical Surface Reconstructlon a ierit ne dischi di tasa dh op era UR das 19 Te Visual Bram ue 29 S Suriace RECOMS UIC MOINS i ee 30 Stee Temple C Usos dien aisles tsi p MOX Uva aay een eee ess 33 10 2D 5D 5urtace Relations BIDS iie n E E pre rop ES Ee per ASENN 33 Cortical Parcellatton Sy stel RE 34 12 Norma
65. c distortion We considered the following tools for creating cortical flat maps in this case study e Computerized Anatomical Reconstruction and Editing Tool Kit CARET http brainmap wustl edu caret e mrUnfold 5 0 http white stanford edu brian mri segmentUnfold htm e BrainVoyager http www brainvoyager de e FreeSurfer http surfer nmr mgh harvard edu We selected Caret to flatten surfaces for two reasons First SureFit was selected for image segmentation and is distributed and supported by the same lab as Caret Thus SureFit is designed to interface seamlessly In fact work is underway to incorporate SureFit into the Caret software suite Secondly since we selected Caret as one of the spatial normalization methods using the same software suite for flattening made for a streamlined evaluation protocol 2 3 Target Atlases Ideally a target atlas will not bias the final solution An ideal template would consist of the average of a large number of MR images that have been registered to within the accuracy of the spatial normalization technique Ashburner and Friston 2000 Talairach Jean Talairach and Pierre Tournoux created the now famous book Co Planar Stereotaxic Atlas of the Human Brain in 1988 Talairach and Tournoux dissected and photographed a post mortem brain of a 60 year old female subject creating a proportional coordinate system often referred to as Talairach space for neurosurgical studies This atlas w
66. caret matrix method euclidean POSTNORM CARET DATA SET OF DISTANCES BETWEEN LANG SITES IN BOTH BRAINS calcdist caret matrix lt as matrix calcdist caret caret bIDs lt langcoord caret Brain ID caret DAB vec DAB caret bIDs calcdist caret matrix caret avgDAB mean caret DAB vec TETTE ae E aE TE AE E HE FE AE FE TE EEE EEE aaa AVG POSTNORM SPM2 DISTANCES ACROSS ALL BRAINS HEFE at E FE OE E TE FE AE AE AE FE AE FE TE TE TE E dg AE E TE TE AE E TE FE TE E E TE dg E E E E E E EE T coord spmimatrix langcoord spm c X yv 2 POSTNORM SPM DATA SET OF LANGUAGE SITE COORD FOR BOTH BRAINS calcdist spm lt dist coord spm matrix method euclidean POSTNORM SPM2 DATA SET OF DISTANCES BETWEEN LANG SITES IN BOTH BRAINS Galbtoodrst sSpmmatrrx lt ds5 matrrxtcalcoorStspm spm bIDs lt langcoord caret Brain ID Spm DAB vec lt DAB Spm oOLIDS calodistspimimatri1x spm avgDAB mean spm DAB vec 120 IEEE EEE HEH EH EE HH HE HEH HEH RE ERE RE ERE ERE E EH EE E RE ERE RE E EE E B E PRESENTING RESULTS t HE HEH HE HH EE ERE ER EE RE ERE RE E EH EE HE EE E RE ERE RE E EH E B E par mfrow c 1 2 final results matrix c expect postnorm avgDAB caret avgDAB spm avgDAB ncol 1 feolnames final results lt g Expecredg CARETM SPMZ rownames final results c Expected Caret SPM2 barplot t final results beside T main 3D Language Site Spread ylim c
67. ck your borders including their orientation 9 4 View normalization results as described on page 62 of the Caret5 User s Manual and Tutorial Version 5 1 April 9 2004 97 10 SPM2 normalization of individual volume to atlas Note This protocol assumes we are using a right handed coordinate system defaults analyze flip 0 in spm defaults m neurological convention right on right The input volumes are Minc files whose X increases from patient left to right and no flipping was done during normalization 10 1 In SPM2 select fMRI Time Series 10 1 1 Change the following defaults Select Defaults Spatial Normalisation Writing Normalised Template bounding box Select Defaults Spatial Normalisation Writing Normalised Voxel size 1 1 1 10 2 Select Normalise Determine parameters and write normalised 10 3 Select Template image T1 mnc 10 4 Select Source image Pxxx Exxxxx Sx 111 mnc 10 5 Select Image to write Pxxx Exxxxx Sx 111 mnc 10 6 At Select Source image subj 2 select Done 10 6 1 Output will be generated after a few minutes and stored in directory as wPxxx Exxxxx Sx lll hdrand wPxxx Exxxxx Sx lll img files 10 7 Select Toolboxes Deformations 10 8 Select Deformations Deformations from sn mat 10 9 Select sn mat file Pxxx Exxxxx Sx 111 sn mat 10 9 1 Output will be generated a few minutes later and stored in directory as y Pxxx Exxxxx Sx lll hdrand y Pxxx Exxxxx Sx 111 img files The indicator for process completion is that
68. ct this change will have on the mapping results We expect that this modification will increase the accuracy of the surface based method by 5 10 66 The metrics used in this study were designed based on the nature of the CSM functional data It would be valuable to evaluate the methods using metrics used to validate other methods For example dispersion metric of selected landmarks overlap percentage and cross comparison of maps would be interesting measures to use to further evaluate surface based and volume based methods Repeating the study using FreeSurfer BrainVoyager and AFNI would provide more insight into how different surface based and volume based methods compare to each other and across categories Evaluating hybrid spatial normalization methods like HAMMER and ROMEO which employ feature based and intensity matching techniques would also be valuable Additionally a relatively recent development is the use of maximization of mutual information MMI registration MMI is a strategy that has proved extremely successful at automatically computing the registration of 3 D multimodal medical images of various organs from the image content Mutual Information MI is a basic concept from information theory that is applied in the context of image registration to measure the amount of information that one image contains about the other The MMI registration criterion postulates that the MI 1s maximal when the images are correctly aligned T
69. d line in the Pxxx directory as follows 11 6 1 Vecwarp matvec in vec input Pxxx xxxxx Sx_111 L full segment_vent_corrx fiducial magctr xxxxx coord gt pxxx fbg coord 11 6 2 Vecwarp apar Pxxx Exxxxx_Sx_111 acpc HEAD input pxxx fbg coord gt pxxx warp coord 11 6 3 Vecwarp matvec out vec input pxxx warp coord gt pxxx acpc coord 11 7 Test that this process worked properly as follows 11 7 1 Open AFNI 11 7 2 Switch to ACPC view 11 7 3 Select Define Datamode 11 7 4 Select Write Anat This writes a Pxxx Exxxxx Sx_111 acpc BRIK file 11 7 5 Quite AFNI 102 11 Create CSM coordinate files continued 11 7 6 Use text editor to create the following spec file BeginHeader Category INDIVIDUAL Date Mon Mar 7 2005 Encoding ASCII Hem flag left Species Human EndHeader volume anatomy file Pxxx Exxxxx Sx llltacpc HEAD PXXX EXXXXX 5x llldTacpc BRIE CLOSET reet 1S PXXX EXEXXXXX EE Pull Segment vent COLrrx XXxXxx LOpo FIDUCTALCOOrd tile pxxxacpce coord 11 7 7 Start Caret and open the spec file Select all 11 7 8 Switch to volume vie and press D C 11 7 9 Switch to Overlay Underlay Volume from the menu at the top of the D C menu 11 7 10 Toggle on the Show Surface Outline checkbox at the lower right 11 7 11 The contour should align perfectly with the volume display If there is a translation or misalignment something has gone wrong and you will need to back track identify and solve the problem before
70. d spherical morphing takes 20 minutes on a Dell Precision 450 dual xeon processors running RedHat 8 Linux 5 25 When finished processing an Align Surface to Standard Orientation dialog will appear along with two bigger dialogs reporting the number of crossovers for the flat and spherical maps More than 10 15 crossovers means there is some concern regarding the quality of the flattening 5 26 If number of crossovers is acceptable click OK to dismiss the statistics dialogs 5 27 If necessary use Viewing Window 2 L and D views to locate the ventral and dorsal ends of the central sulcus 5 28 In the main caret window flat map click on the ventral tip of the central sulcus and shift click on the dorsal tip 5 29 On the Align Surface to Standard Orientation dialog check the Align Sphere box and click Apply 5 30 File Save Data File Coord file and replace FLAT CYCLES OVERLAP SMOOTH in the filename with FLAT Cartesian Change the coord fram to Cartesian Standard Save the file 5 31 File Save Data File Coord file and select SPHERICAL CYCLE4 from the coord file drop down menu Replace CYCLE4 in the filename with Std Change the Coord Frame to Spherical Standard Change the orientation to Left Posterior Inferior 5 32 File Save Data File Latitude Longitude and save the lat lon coordinates generated during alignment to Pxxx_Exxxxx_Sx_111 L xxxxx latlon 84 5 Surface Flattening continued 5 33 Close the Align Surface t
71. e Handle Count results in 0 2 handles Save Edits and note the final corrected version number You are now ready to perform the final segmentation and prepare for flattening 4 15 Final Segmentation Use Segmentation in Vol 2 Fill Ventricles No Generate Surface deselect Correct Errors and select Identify Sulci Leave Keep intermediate files unselected Prepare to flatten Yes 4 16 When finished exit SureFit 82 5 Surface Flattening 5 1 Change to SURFACES and make a back up copy of the spec file cp Pxxx Exxxxx Sx_111 L full segment_vent_corrx Surface xxxxx spec Pxxx corrx Surface xxxxxx spec bak 5 2 Open caret5 and select spec file 5 3 For species select Human leave Space field blank for Category select INDIVIDUAL 5 4 select all and load 5 5 Surface Flatten full or partial hemisphere 5 6 Flattening type Full hemisphere ellipsoid and morph sphere 5 7 Accept Surfaces defaults fiducial and ellipsoidal 5 8 Accept AC position offset defaults 5 9 Choose Human left standard cuts and make sure Smooth Fiducial Medial Wall is selected select OK 5 10 Pull Continue Flattening full hemisphere dialog to the side 5 11 Select Window Viewing Window 2 5 12 Switch window 2 to the INFLATED view and resize the window larger 5 13 Press M on window 2 to switch to a medial view 5 14 Select XY on the drop down menu to switch the rotation axis to XY 5 15 Rotate the surface so the the ventral side is a
72. e become organized to process visual information There are also demonstrable differences found across the brains of deaf people who use sign language and those who do not These structural and functional differences are presumably due to differing language experiences Another example of plasticity akin to the deaf reorganizing temporal cortex for visual processing is blind subjects visual cortex reorganizing for language processing Blind subjects asked to generate verbs in response to heard nouns showed changes as measured by fMRI in the visual cortex Responses were greater and broader in early blind subjects than in late blind subjects Burton 2003 The functional data used in our research is cortical stimulation mapping for language localization described in Section 1 5 Figure 2 shows cortical stimulation sites identified as statistically significant for naming errors are located in different areas of subjects left hemispheres This example of the same function being located in clearly different parts of the brain demonstrates the marked individual functional variability in cortical locations essential for language production Figure 2 Functional Variation surface reconstructions of four subject s left hemispheres The green spheres represent sites that interrupted language production when electrically stimulated during awake neurosurgery Note how sites responsible for the same function appear on different areas of the cor
73. e deformation and are widely used in medical applications where the structures of interest are either bone or enclosed with bone and are commonly used to register head and brain images In the case of intersubject registration however rigid body normalization does not provide enough DOF for adequate intersubject registration Some registration algorithms increase the DOF by allowing for anisotropic scaling giving nine DOF and skews giving 12 DOF This type of transformation is referred to as affine and can be described in matrix form Also an affine transformation preserves parallel lines A rigid body transformation can be considered a special case of the affine in which scaling values are unity and skews equal zero Hill et al 2000 While Heller er al did not find significant differences between an affine a rigid and three non affine normalization methods when evaluating local measures based on matching of cortical sulci they did find that for global measures the quality of the registration is directly related to the transformation s DOF Heller et al 2003 Collins and Evans compared rigid and non affine normalization methods In this study the rigid method revealed problems in maintaining accurate global head shape and shape of internal structures like the ventricles as well as an error rate more than 50 higher than the non affine method Collins and Evans 1997 Crivello s comparison of one simple affine and three non affine nor
74. e image and artifacts especially in the left temporal 30 lobe our primary region of interest Working with a SureFit expert we were able to screen out images that would require substantial manual error correction due to poor image quality In one instance P62 non uniformity of intensity was an issue Using FSL s fast algorithm http www fmrib ox ac uk fsl fast significantly improved the quality of the image making it possible to include the image in the data set The final level of screening eliminated subjects with large lesions Step 2 Create Surface Reconstruction As discussed in Section 2 1 we selected SureFit to create surface reconstructions of the fourth cortical layer of the left hemisphere of each subject s brain Figure 8 contains three of the 11 surface reconstructions segmented for this study and the surface reconstruction of the target atlas colin27 Figure 8 Surface reconstructions of four left hemispheres created using SureFit 3l Prior to launching the automated segmentation process the MRI volume was resampled to 1 mm cubic voxels and cropped to included only the left hemisphere The segmentation generated a cortical surface reconstruction in approximately 1 2 hours using a Dell Dimension dual 450Mhz processor running Debian Linux Error detection and correction involved automatic correction and interactive editing Topological errors called handles in the initial segmentation are typically at
75. e input for the surface based method is the surface reconstruction and the coordinate file At this point the automatic surface based normalization takes about the same time as the volume based method 15 20 minutes If a normalized set of deformed CSM coordinates is the only desired result from the anatomical normalization process then the volume based method is less expensive and will provide an overall accuracy of approximately 78 If visualization of the results is desirable then the surface based method is superior to the volume based method which is not designed for visualization of cortical site data SPM2 is used to visualized fMRI however Without creating the surface reconstructions required for Caret we would not have been able to localize the notable variation between the subjects average surface and the target atlas in the superior temporal gyrus or assess the bias introduced by the colin27 atlas atypical folding patterns in the supramarginal gyrus and ventral part of the central sulcus Additionally the visualization of the CSM mappings was critical to assessing method accuracy Knowledge of the accuracy of a given method is key to researchers choosing the best spatial normalization method for their work The challenge of validating the volume based method is discussed by Crum He advocates for registration tools that monitor their own performance and estimate correspondence error with minimal intervention Crum et al 2003
76. e3 Error Total Error 54 0 0 5 5 55 1 0 8 9 58 1 0 4 5 60 2 1 5 8 61 3 0 3 6 62 3 0 p 5 63 6 0 2 8 117 0 0 0 0 164 2 0 3 7 170 3 0 p 5 176 3 0 l 4 Totals 24 1 37 62 49 Figure 23 Caret mapping of 21 language sites viewed on colin27 inflated surface reconstruction Figure 24 SPM2 mapping of 21 language sites viewed on colin27 inflated surface reconstruction 50 Error Rate Analysis by CPS Parcel The middle part of the superior temporal gyrus MSTG had the most assigned CSM sites accounting for over 12 24 of 198 sites of all CSM sites mapped The average error rate in this parcel as measured by averaging the sum of the Caret error rate and the SPM2 error rate was also high 54 2 of MSTG sites were incorrectly mapped Other parcels with 7 or more assigned sites having an error rate of 50 or more were the posterior part of the supramarginal gyrus PSMG and the ventral part of the precentral gyrus VPrG Number of CSM Sites Per Parcel gm Num Sites CSM Sites e 46 49 PD A 46 OG GO GGL P D D ZS E ZS A ES a S gr EES Pu Soc Parcel Name Figure 25 Summary of 198 CSM sites included in this study broken down by CPS parcel Error Rate by CPS Parcel 120 0 100 0 80 0 60 0 mw Avg Errors Parcel 40 096 Percentage of Errors 20 0 0 0 O O C O O O SS eq uw o e V OQ OG ai o 06 9 0 0 06 0 Sg eua og CES Parcel Name
77. econstructions Red dashed line traces the lateral fissure Notice how the posterior end of this landmark circled in red differs across brains Ono s study provides a broad analysis of cortical folding patterns A localized example of variation is demonstrated in figure 1 In this figure there are two volume reconstructions created from Magnetic Resonance Images MRI using SKANDHA a 3 D visualization software tool Prothero 1995 The posterior tip of the lateral fissure is circled in red The difference between individual cortical folding patterns within the circled region is clear even to an untrained observer This degree of gross cortical structural variation between any two individuals makes it difficult to accurately compare cohorts 1 3 Functional Variation The anatomy of the brain houses sensory motor and cognitive functions While language related functions were the first to be ascribed to a specific location in the human brain Broca 1861 there is much more validation of and consensus around the anatomical location for sensory and motor functions A classical model of language organization based on data from aphasic patients with brain lesions was popularized during the late 19 century and remains in common use Binder et al 1997 In its most general form this model defines a frontal expressive area for planning and executing speech and writing movements named after Broca and a posterior receptive area for anal
78. ect File Save Window as Image DX X X CSM flat jpg 11 3 14 Switch to inflated and fiducial views and save those captures if desired 11 4 Align volume to AC PC space 11 4 1 11 4 2 11 4 3 11 4 4 11 4 5 11 4 6 11 4 7 11 4 8 11 4 9 To create AFNI files HEAD and BRIK with the minc start variable call 3dMINCtoAFNI Pxxx Exxxxx Sx_111 mnc at the command line Call 3drefit markers Pxx Exxxxx Sx 111 HEAD before starting AFNI Start AFNI To view axial sagittal and coronal views click on image buttons Select DEFINE MARKERS Check allow edits box Set AC superior edge by finding it in the slice windows Once the cross hairs are aligned with the superior edge select Set AC Superior Edge Set AC posterior margin by moving 1 slice posterior and 1 slice inferior of the AC superior edge Select Set AC Posterior Margin Set inferior edge of the PC by finding it in the axial slice window Once the cross hairs are aligned with the inferior edge select Set PC Inferior Edge 100 11 Create CSM coordinate files continued 11 4 10 Select 2 mid sagittal points that are in the same sagittal plane You want these points to contain CSF After finding each of the two desired locations select Set Sagittal Point 11 4 11 Check that the sagittal points are acceptable by selecting the Quality button If no warning you can proceed Otherwise reselect the 2 sagittal points and recheck quality 11 4 12 Select Transform Data button
79. elect the Core6 landmarks required to constrain the registration as discussed in Section 2 4 1 To delineate landmarks traces along the sulci fundi inflated surfaces of the individual hemisphere and the corresponding colin27 atlas hemisphere were viewed side by side Endpoints for each landmark trace were drawn as prescribed by the Core6 protocol Complete landmark traces were then drawn between the endpoints on the flat map using the visualization of cortical folding to determine the trajectory of the fundus see figures 13 and 14 Landmark contours were projected onto the surface and saved as a border projection file in a barycentric format The file was then mapped to the spherical standard surface that was used for registration Spherical registration was started using the Deform Spherical Map function with the deformation selection Deform Individual to Atlas and Deform Atlas to Individual Drawing landmarks for one subject took 15 30 minutes The automatic normalization process took approximately 15 minutes per subject on a Dell Dimension dual 450Mhz processor running Debian Linux 38 Figure 13 Inflated left hemisphere surface reconstruction with three of the Core6 landmark traces The yellow trace is along the fundus of the central sulcus turquoise trace is along the lateral fissure fundus and pink trace is along the tip of the superior temporal gyrus Figure 14 Spherical surface used for registration to atlas
80. ential for language function These sites are considered essential for language function because 1 resecting tissue close to such areas usually results in postoperative aphasis i1 avoiding them by 1 5 to 2 cm avoids such language deficits and 111 all aphasic syndromes include anomia Modayur er al 1997 In addition the patient s responses and markers indicating when and where stimulation occurred is recorded on audio tape and used to later check the results to determine if the sites that were identified as essential for language function are statistically significant for naming errors as determined by the subject s responses with and without stimulation 14 Figure 3 Surgical photograph taken during preparation for a left hemisphere temporal lobe resection The numbered tags identify sites that have been electrically stimulated referred to as cortical stimulation Sites A visual comparison approach is used to transpose the location of the numbered tags as seen in figure 3 to a 3D volume reconstruction of the cortical surface including veins and arteries The interactive mapping process is facilitated by the Skandha4 software package As shown in figure 4 the language mapping module graphic user interface GUI includes the intraoperative photo the MR volume data a 3D rendered image of the brain and a palette of numbers The neuroanatomist expert determining the localization of the sites is blind to whether or not sites had naming error
81. ese sulci tend to have more stable sulci patterns compared to other landmarks The central sulcus located on the lateral surface of the frontal lobe is the most important and constant landmark on the convexity of the brain However even this sulcus has notable variation upon visual inspection of the shape of its inferior end Ono found three different types of shapes in the specimen brains In the right hemisphere 52 of the inferior ends were straight 28 had a Y and 20 had a T shape In the left hemisphere it was found that 80 of the inferior ends were straight and 20 were T shaped as outlined in table 1 Table 1 Incidence rates of inferior end of central sulcus patterns as determined by Ono et al INFERIOR END SHAPE LEFT HEMISPHERE RIGHT HEMISPHERE Straight Table 2 Incidence rates of superior temporal sulcus patterns found in 25 human autopsy specimen brains PATTERN LEFT HEMISPHERE RIGHT HEMISPHERE Interrupted w 2 segments 32 48 Interrupted w 3 segments 16 Greater variability was found on the lateral surface in the superior temporal sulcus In table 2 there are four patterns and the occurrence rate of those patterns Notice how the patterns in this case are much less predictable than the shape of the inferior end of the central sulcus POSTERIC 0 AN HE i AM or Perm OMIM SC dad i G G i akat did nila ED lai anian TT K Figure 1 Left hemisphere of two different subject s volume r
82. est p lt 0 05 was used to compare each subject s baseline performance derived from the naming error rate in each control trial associated with the site regardless of target and performance under stimulation at that site This definition of baseline restricted to the controls associated with a certain site was established to eliminate variation in performance due to fatigue inattention and other physical factors experienced by the subject during the procedure The p value represents the reliability that an error was observed under stimulation relative to the unstimulated baseline for each individual site Corina et al 2005 36 Table 4 Summary of statistically significant language site anatomical localization ID CSM Location ID CSM Z Location P54 20 PSTG P61 30 PSMG AnG P54 30 PSTG P62 33 OpIFG P54 35 PSTG P62 35 TriFG P54 36 PSMG P63 25 MSTG P55 41 MSTG P117 21 ASMG P58 32 PMTG P117 33 PSMG P60 29 MSTG P164 35 VPRG Peo 31 PMTG P164 40 ASMG P61 25 MSTG P170 26 MSTG P61 28 PMTG P176 28 MSTG Generating pre normalization coordinates that could be compared across subjects required shifting the individual volume images into a common grid i e standard voxel size origin and orientation Without a common grid we could not reasonably establish baseline distances bet
83. face draw the central sulcus by dragging the mouse with the left mouse button depressed When you get to the dorsal extent delimited by your green ID node shift click the left mouse button to terminate the border 8 19 16 On the Draw Border dialog select LANDMARK SylvianFissure from the Name menu and click Apply 8 19 17 Starting at the dorsal tip of the Sylvian fissure just across from the origin scale bar near the center of the surface draw the Sylvian fissure by dragging the mouse with the left mouse button depressed When you get to the ventral extent delimited by your green ID node shift click the left mouse button to terminate the border 8 19 18 On the Draw Border dialog select LANDMARK SF STSant from the Name menu and click Apply 8 19 19 Starting at the dorsal tip of the superior temporal gyrus closer to the center of the surface trace along the white line until you reach the end point then shift click 8 19 20 If you re not happy with a border select Layers Border Delete border with mouse as needed to delete a bad border and redraw 93 8 Drawings borders in Caret for surface normalization continued 8 19 21 Click on a node at the temporal pole corresponding to the end point of the magenta colored SF STSant border 8 20 With ID terminal points for each border on the inflated map we can draw the borders on the flat map Aim for the fundus dark line of the central sulcus and sylvian fissure as it appea
84. han other organs in the body For example the human lung has a single primary function respiration On the other hand brain function includes regulating all bodily functions language the five senses and conscious thought among others When looking at structural features like shape surface and borders of the lung versus the brain we again are reminded of the brain s complexity In Gray s Anatomy it requires four times the amount of visual and spatial description to characterize gross brain anatomy compared to that of the lung The cerebral cortex in particular reveals the brain s structural complexity The sulci concavities and gyri convexities as viewed on the cortical surface serve as key landmarks to neuroscientists The Atlas of Cerebral Sulci by Ono et al is a reference book documenting anatomical variation of the cerebral sulci as a step toward describing and categorizing the highly varied structural patterns of the cortical surface Ono compared sulcal patterns of 25 autopsy human specimen brains examined for anatomical variation and consistency in location shape size dimensions and relationship to parenchymal structures Ono et al 1990 Ono analyzed a total of 28 sulci 15 large main sulci six short main sulci and seven others Six of the large main sulci are key landmarks They include central sulcus lateral fissure AKA Sylvan fissure collateral sulcus callosal sulcus calcarine sulcus and parieto occipital sulcus Th
85. he MMI criterion is volume based uses a histogram instead of intensity matching and does not impose limiting assumptions on the nature of the relationship between corresponding voxel intensities MMI has been shown to be a very general and powerful criterion that can be applied automatically and reliably without prior segmentation or preprocessing on a variety of applications Maes et al 2003 It would be interesting to compare the results of other surface based and volume based spatial normalization methods to a MMI method like that employed by Rueckert or D Agostino 67 List of References Ashburner J Csernansky JG Davatzikos C et al Computer assisted imaging to assess brain structure in healthy and diseased brains Lancet Neurol 2003 2 2 79 88 Ashburner J Friston KJ Voxel based morphometry the methods Neuroimage 2000 11 805 821 Binder JR Frost JA Hammeke TA et al Human Brain Language Areas Identified by Functional Magnetic Resonance Imaging J of Neurosci 1997 17 1 353 362 Brain Innovation lt http www brainvoyager com gt Accessed December 2003 Bransford JD Brown AL Cocking RR Eds How People Learn Brain Mind Experience and School National Academies Press 1999 lt http www nap edu openbook 0309065577 html gt Brett M The MNI brain and the Talairach atlas MRC Cognition and Brain Sciences Unit lt http www mrc cbu cam ac uk Imaging mnispace html gt Accessed Dec 2003 Brink
86. hod would 1 minimize variation as measured by spread reduction between cortical language sites across subjects while also 11 preserving anatomical localization of sites Evaluation technique Eleven MR image volumes and corresponding CSM site coordinates were selected Images were segmented to create left hemisphere surface reconstruction for each patient Individual surfaces were registered to the colin27 human brain atlas using each method Deformation parameters from each method were applied to CSM coordinates to obtain post normalization coordinates in 2D space and 3D ICBM152 space Accuracy metrics were calculated 1 as mean distance between language sites across subjects in both 2D and 3D space and i1 by visual inspection of post normalization site locations on a 2D map Results Globally we found no statistically significant difference between CARET surface based method and SPM2 volume based method as measured by both spread reduction and anatomical localization Local analysis showed that more than twenty percent of total mapping errors were mapped incorrectly by both methods There was a statistically significant difference between Caret and SPM2 mapping of type 2 errors TABLE OF CONTENTS Page IB COMO EE ii IE e TOble Suc dum NM mE ANDE MEM UEM IM UE iv wise 1 INT OCUCHION ET OO l J Resistration of Medical Mages a a caa a eee coe ido te ia cote o race I 1 2 Anatomica E EEN 2 BSE UMC tonat DEE 4 1 4 Survey of Spatial Normaliz
87. how we cannot expect to completely normalize folding patterns across individuals It also highlights the bias introduced by a single brain atlas In this study we hypothesize that areas of important variability between source and target are a key cause of at least 20 of the total anatomical localization errors Possible solutions to the problem of important anatomical variation between source and target include creating a probabilistic atlas of epilepsy subjects perhaps using the same 11 subjects This atlas would incorporate the average sulcal shape of the subjects presumably resulting in better anatomical alignment and more accurate normalization Extrapolation of these findings to the normal brain would require a transform of the functional data to a normal subject atlas An atlas of normal brains like the PALS B12 or ICBM atlases would be a preferable target What currently prevents this technique from being implemented is the limited functionality available to map CSM functional data to these atlases There currently exists functionality to map fMRI data to these atlases as MRI is the most commonly mapped functional data However CSM site data is relatively rare in the neuroscience community and therefore is not accommodated for as broadly as is fMRI Additionally using a probabilistic atlas introduces a problem in reliably measuring post normalization surface distances that would need to be addressed and is discussed in Section 6 62
88. hzatiom N SSe EE TA 37 13 Core6 Landmarks Inflated Surface 38 14 Core6 Landmarks Spherical Surface essen 38 I5 4D ISTAHC CI le Ul Guetta Deen Lean aia a ALIS E 4 16 reene 43 17 DLL TA Oa E S OS wears uo e missa suo ahis IN ae pti tam dau aera oes 43 5 Ee a a st ntn th eat te ent etn eat ae est E ae ec E M Erde 44 I9 Type Ee EE 45 EN D i E 45 2 2D Spread N a ee E 47 RST R Spread REGUCH OMS ett 47 25 are Eanguadee Site E er 49 24 5PM2 Language Site MapDIp 50 ee ee 49 25 Number of CSM Sites per Parcela itu uve V EUM UHR RM EU UR NNUS 50 26 xor Rate Dy C e RE 50 23 FETT Ol ey ey KEEN 5 lii LIST OF FIGURES continued Figure Number Page 25 E Ee EE 56 29 Refined 3D Spread AR COUCTIOM sssini cnc ipu Pen r UU HN MN essei ep US 56 DUC COL V EOS petra ctl Ant rta tai tnit aie cipit enitn pani cto tni 58 S1 C PS Parcel DIAS E 59 224 ee 60 23 5ulcal Ditfetence E E 60 34 Language Site Mapping Summary sese 64 1V LIST OF TABLES Table Number Page I Central Serge 3 2 S perior Lemporal Sulcus Patterns sco e eee en nest eise RN duis m ME Mi cem dd UE 3 Eeer EE Mu s LS ones 28 A Statistically Significant Language Sites e oo oe nore a qe ee ae LOG ce ede te eie m edu eere 36 5 oubjecbourtdces and Volume ATCAS 3s eere et so ee s ER b co iare dido sees eiit 42 6 Results SUNMA suse tM MEM E E MEE LE EDI NE DU 46 EE 48 S SPM Error SUMMA E 48 Z Same Eror KRU Tee 52 T0 Umgue Caret oh e 53 Li Ugue SEM E OLS EE 54 12 Lano
89. ial normalization methods that have been developed and used over the years Others include deformable templates using large deformation kinematics Christenson et al 1996 elastic deformation algorithm Gee et al 1993 intersubject averaging and change distribution analysis Fox et al 1988 unified framework for 10 boundary finding in a Bayesian formulation Wang et al 2000 statistical and geometric image matching Gee et al 1994 automated image registration AIR Woods et al 1998 octree spatial normalization OSN Kochunuv ef a 1999 automatic non linear image matching and anatomical labeling ANIMAL Collins and Evans 1997 analysis for functional neuro images AFNI Cox 1996 and maximization of mutual information MMI Rueckert er al 2001 D Agostino et al 2002 Surface based non linear spatial normalization methods include STAR hybrid surface models Thompson and Toga 1996 deformable surface algorithm Davatzikos and Bryan 1996 generalized Dirichlet solution for mapping brain manifolds Joshi et al 1995 thin plate splines Bookstein 1989 unified non rigid feature registration Chui et al 2003 computerized anatomical reconstruction and editing tool kit Caret Van Essen et al 2001 Freesurfer Fischl et al 1999 BrainVoyager Kiebel Goebel and Friston 1999 iconic features PASHA Cachier et al 2002 and active ribbons Bizais 1997 There are also spatial normalization methods
90. ightly posterior to the gyral inflation that is just posterior to the postcentral sulcus Anteriorly the landmark extends almost to the anterior and ventral limit of the primary fundus 10 mm dorsal to the ventral margin of the frontal lobe For more detail see Spherical Registration to Atlas Core6 landmark set link at http oub download brainmap wustl edu pub donna WASHINGTON 200503 p117 html 8 17 Identify Superior Temporal Gyrus extent as follows 8 17 1 Use Toolbar L in the inflated view and again rotate about Z if needed to AC PC align the individual s surface to match Colin s alignment 8 17 2 Click on a node along the superior temporal gyrus lower edge of the Sylvian fissure directly below the node ID s for the ventral tip of the central sulcus 8 17 3 Click on a node at the temporal pole corresponding to the end point of the magenta colored SF STSant border 8 18 With ID terminal points for each border on the inflated map we can draw the borders on the flat map Aim for the fundus dark line of the central sulcus and sylvian fissure as it appears on the flat map s folding pattern even if your ID marks miss the fundus The SF STSant landmark is a gyrus so aim for the whitline The green ID nodes show you where to start and stop drawing Draw borders as follows 8 18 1 Layers Border Draw Border 8 18 2 Name LANDMARK CentralSulcus Resampling 4 0 Apply 9 8 Drawings borders in Caret for surface normalizat
91. in and smaller objects will disappear from Vol 2 4 12 As you correct errors using these tools you will want to perform two task periodically 4 12 1 4 12 2 After making a correction check whether or not you reduced the number of handles by selecting Update Handle Count If you have successfully corrected a handle save the patched volume by selecting Save Edits In the pop up file selection window the default file name will have patch appended to the current Vol 2 file name Instead of using this file name delete this appendage and revise the name to corr2 to identify the second corrected volume and repeat this protocol until all errors are corrected 81 4 Interactive error correction continued 4 13 In case of mistakes the Undo button can be used to recover the most recent step taken Changes made at earlier steps are recoverable if intermediate volumes have been saved in which the partially patched volume can be reloaded 4 14 To determine if the surface reconstruction is satisfactory after patching generate a new surface as follows 4 14 1 4 14 2 In the Run SureFit window select Use Segmentation Loaded in Vol 2 and Skip Error Correction Unlike the Update Handle Count function which operates on unsaved patches generating a new surface uses the most recently saved version of the segmentation Check the fiducial and inflated surfaces for irregularities If the surfaces look clean and the Updat
92. in mapping Proc AMIA Annu Fall Symp 1997 429 433 Ojemann G Ojemann J Lettich E et al Cortical language localization in left dominant hemisphere J Neurosurg 1989 71 316 326 Ono M Kubik S Abernathy CD Atlas of Cerebral Sulci New York 1990 Penny W http www fil ion bpmf ac uk spm software spm2 Accessed Jan 2005 Roland PE Geyer S Amunts K et al Cytoarchitectural maps of the human brain in standard anatomical space Human Brain Mapping 1997 5 222 227 Rueckert D Frangi AF Scnabel JA Automatic construction of 3D statistical deformation models using nonrigid registration Lecture Notes in Computer Science Proc Medical Image Computing and Computer Assisted Intervention 2001 Oct 2208 77 94 Salmond CH Ashburner J Vargha Khadem F et al The precision of anatomical normalization in the medial temporal lobe using spatial basis functions Neuroimage 2002 Sep 17 1 507 12 Senda M Ishi K Oda K et al Influence of the ANOVA design and anatomical standardization on statistical mapping for PET activation Neuorimage 1998 8 283 301 Segmentation and Flattening Software Wandell Lab lt http white stanford edu brian mri segmentUnfold htm gt Accessed Dec 2003 Statistical Parametric Mapping Wellcome Department of Imaging Neuroscience lt http www fil 10n ucl ac uk spm Accessed Dec 2003 Sugiura M Kawashima R Sadato N et al Anatomica validation of spatial normalization methods for PET J Nuc
93. indow is set to the FIDUCIAL surface Surface Transform Translate and enter the Xmin Y min Zmin values Translate X 50 Translate Y 23 Translate Z 91 Caution mincinfo lists the origin in z y x order not x y z So take care entering the parameters below Surface Transform Translate Translate X 117 7 Translate Y 114 5 Translate Z 107 6 86 6 Translate cropped volume to magnet coordinate space continued 6 4 7 File Save Data File Coord File append magctr after fiducial in the filename In the comment field enter Translated 50 23 91 Xmin Ymin Zmin then 117 7 114 5 107 6 Minc start Leave the coord frame Native but change the orientation to Left Posterior Inferior Save 6 4 8 Press Toolbar Spec and locate the entry corresponding to the original fiducial coord file 1 e without magctr in the name press X to remove this file from the spec file so you don t select the wrong fiducial surface when mapping foci or registering the surface to colin 6 4 9 Also click X to remove these spec file entries RAW all SPHERICAL except Std ELLIPSOID COMP MED WALL all flat except Cartesian all border files TEMPLATE CUTS bordercolor 6 5 Check to make sure the translated surface aligns with the volume as follows 6 5 1 Use text editor to create this spec file BeginHeader Category INDIVIDUAL Date Mon Mar 7 2005 Encoding ASCII Hem flag left Species Human EndHeader CLOSEDtopo file
94. inear combination of smooth spatial basis functions Ashburner and Friston 1999 The nonlinear registration involves estimating the coefficients of the basis functions that minimize the residual squared difference between the image and the template while simultaneously maximizing the smoothness of the deformations This step has been improved in that the bending of energy of the warps is used to regularize the procedure rather than membrane energy This model seems to produce more realistic looking distortions It is worth noting that this method of spatial normalization corrects for global brain shape differences but does not attempt to match other cortical features Ashburner and Friston 2000 28 Section 3 Methods Subjects The subjects were 11 patients 5 female 6 male age range 23 52 years undergoing resection treatment at the University of Washington Medical Center for chronic epilepsy n 11 Seven patients were right handed All cortical stimulation occurred in the subject s left hemisphere which was identified as the subject s language dominant hemisphere in all subjects determined by pre surgery WADA testing Corina et al 2005 Subject demographics are summarized in table 3 Table 3 Subjects gender age handedness and verbal IQ VIQ 54 gt 23 38 N OO ON bech Evaluation technique protocol To test our hypothesis we developed a six step evaluation protocol 1 select MRI volumes 2 create su
95. ion continued 8 18 3 Click on a node at the temporal pole corresponding to the end point of the magenta colored SF STSant border 8 19 With ID terminal points for each border on the inflated map we can draw the borders on the flat map Aim for the fundus dark line of the central sulcus and sylvian fissure as it appears on the flat map s folding pattern even if your ID marks miss the fundus The SF STSant landmark is a gyrus so aim for the whitline The green ID nodes show you where to start and stop drawing Draw borders as follows 8 19 1 Layers Border Draw Border 8 19 2 Name LANDMARK CentralSulcus Resampling 4 0 Apply 8 19 3 Starting at the ventral tip of the central sulcus near the origin scale bar near the center of the surface draw the central sulcus by dragging the mouse with the left mouse button depressed When you get to the dorsal extent delimited by your green ID node shift click the left mouse button to terminate the border 8 19 4 On the Draw Border dialog select LANDMARK SylvianFissure from the Name menu and click Apply 8 19 5 Starting at the dorsal tip of the Sylvian fissure just across from the origin scale bar near the center of the surface draw the Sylvian fissure by dragging the mouse with the left mouse button depressed When you get to the ventral extent delimited by your green ID node shift click the left mouse button to terminate the border 8 19 6 On the Draw Border dialog select LAN
96. ioning Probabilistic atlases like the ICBM Tissue Probabilistic Atlas and Lobular Probabilistic Atlas proceed as follows e Classify the desired components tissue type or lobe type in these cases e Average the separate components across the subjects to create probability fields for each component that represent the likelihood of finding each component at a specified position for an individual brain that has been linearly aligned to the atlas space Toga and Mazziotta 2000 24 PALS B12 Atlas The Population Average Landmark and Surface based atlas PALS B12 is a new electronic atlas developed at the Washington University Van Essen Lab Designed for brain mapping analysis it is derived from the MRI volumes of 12 normal young adults and includes both volume based MRI and surface based representations of the cortical shape The population average and individual subject representations were created using Caret a surface based method of spatial normalization discussed in Section 2 4 The atlas includes sulcal depth maps as a standard shape representation and depth difference maps can be used to view differences between individuals and across populations The atlas also includes probabilistic representations of the population average surface and volume Van Essen 2005 This atlas was designed specifically to avoid the inevitable bias introduced when using a single brain atlas as a target A multi fiducial mapping method is introduced th
97. ith the final segmentation Those options include 3 3 2 1 Some images may have non uniform intensity levels For example in three images we processed the occipital lobe intensities tended to be higher so sulci fused over The temporal lobe intensities however tended to be lower so the anterior medial portions of it often fade out of the segmentation Where bias correction was needed we used FSL s best and fast applications as follows Minc2Analyze Pxxx Exxxx Sx_111 L full sMRI mnc bet Pxx Exxxx Sx_111 L full ssMRI Pxx Exxxx Sx_111 L full sMRI bet 0 1 g 0 fast t1 c 3 n v5 1 500 or Pxx Exxxx Sx_111 L full sMRI_bet hdr Analyze2Minc Pxx Exxxx Sx 111 L full sMRI bet restore hdr mv Pxx Exxxx_Sx_111 L fullssMRI bet restore mnc Pxx Exxxx Sx 111 bet bc L full sMRI mnc 3 3 2 2 If this bias correction is performed step 2 Volume Preparation needs to be re done before moving onto final segmentation 78 3 Segmentation surface generation and automated error correction continued 3 4 22 3 6 3 7 3 3 3 Some images may result in ventricle filling problems 3 3 3 1 Ventricle filling problems typically will require a customized work around solution For one of the instances of this problem type we used a VElab utility called VolMorphOps that erodes or dilates the input volume and another utility CombineVols to perform a logical or operation on two volumes For this type of problem it is best to consult with the anal
98. l Med 1999 40 317 322 Prothero JS Skandha4 An Slisp based Interactive Raster Graphics Toolkit lt hittp sig biostr washington edu projects skandha4 tr skandha4 gt 1995 Thirion JP Image matching as a diffusion process an analogy with Maxwell demons Med Image Anal 1998 Sep 2 3 243 260 71 Thompson PM Toga AW A surface based technique for warping three dimensional images of the brain IEEE Transactions on Medical Imaging 1996 15 4 402 417 Thompson PM Cannon TD Narr KL et al Genetic influences on brain structures Nat Neurosci 2001b 4 12 1253 1258 Toga AW Mazziotta JC Brain Mapping The Systems Academic Press San Diego 2000 Toga AW Thompson PM Maps of the Brain New Anat 2001 265 37 53 Toga AW Thompson PM Temporal dynamics of brain anatomy Annu Rev Biomed Eng 2003 5 119 145 Van Essen DC Drury HA Dickson J Harwell J et al An Integrated Software Suite for Surface based Analysis of Cerebral Cortex JAMIA 2001 8 443 459 Van Essen DC Windows on the brain The emerging role of atlases and databases in neuroscience Curr Op Neurobiol 2002 12 574 579 Van Essen DC A population average landmark and surface based atlas of human cerebral cortex Draft 2005 Wang Y Staib LH Boundary finding with prior shape and smoothness models IEEE Transacations on Pattern Analysis and Machine Intelligence 2000 July 22 7 738 743 Wernicke C 1874 Der aphasische Symptomenkomplex
99. ld have preferred a population atlas as our target like PALS 12B because of the inherent structural bias introduced to normalization by any single brain atlas We expect that using a probabilistic atlas would significantly increase the anatomical preservation accuracy of both methods The problem of multi fiducial mapping for CSM sites could be circumvented by using individual deformed files to assign nodes to the CSM sites that could then be viewed on a variety of substrates e g colin27 PALS 12B average subject surfaces etc With this type of visualization we would create a zone for each site that would capture the average location of a given CSM site across subjects and likelihood as to where any given site would fall within this zone This approach would require modification to the spread reduction calculation To achieve statistical significance of p lt 05and 80 power we would want to increase the number of subjects to at least 60 and preferably 100 or more Functional MRI data has been collected on many of the subjects in the CSM database It would be interesting to repeat this study replacing the CSM data with the MRI data using multi fiducial mapping to view results This study could serve as validation of both methods and further contribute to an understanding of what accuracy can be expected when using each method We plan to repeat the surface based method using the recently modified Core6 landmarks to analyze the impa
100. ley JF Rosse C Imaging and the Human Brain Project A Review Methods Inf Med 2002 41 245 60 Brain and Anatomy Function Traumatic Brain Injury Resource Guide Centre for Neuro Skills lt http www neuroskills com edu ceufunction7 shtml gt Accessed Dec 2003 Bookstein FL Principal warps thin plate splines and the decomposition of deformations IEEE Transactions on Pattern Analysis and Machine Intelligence 1989 11 6 567 585 Burton H Visual cortex activity in early and late blind people J Neurosci 2003 23 4005 4011 Cachier P Bardinet E Dormont D Pennec X Ayache N Iconic feature based nonrigid registration the PASHA algorithm Computer Vision and Understanding 2003 89 272 298 Christensen GE Rabbit RD Miller MI Deformable Template Using Large Deformation Kinetics IEEE Transactions on Image Processing 1996 5 10 1435 1447 Chui H Win L Schultz R et al A unified non rigid feature registration method for brain mapping Med Imag Anal 2003 Jun 7 2 113 130 68 Collins DL Evans AC ANIMAL Validation and Application of Nonlinear Registration based Segmentation International Journal of Pattern Recognition and Artificial Intelligence 1997 11 8 1271 1294 Corina DP Gibson EK Martin R et al Dissociation of action and object naming evidence from cortical stimulation mapping Hum Brain Mapping 2005 24 1 10 Cox RW AFNI Software for analysis and visualization of functional magnetic resonance neu
101. library help scatterplot3d itt HH HH HH HHH HH HH HO EE HOE EE T PRENORM LANGUAGE SITE COORD itt HH Ht Ht HH HH OH HO EEE HOE EE T langcoord prenorm pre dat 0 temp matrix langcoord prenorm length dim pre dat 1 forti mpm he lengra 4 if pre dat i CSM Region 1 temprow pre dat i temp matrix lt rbind temprow langcoord prenorm langcoord prenorm lt temp matrix END OF FOR LOOP HEE HHH EE HH HEE HH HEE HH HEE HH HE EF CARET LANGUAGE SITE COORD HEE HEHEHE EE HH HEE HH HEE HH HEE HH HE EF TangGCoOOord caren lt pest dat E temp matrix langcoord caret Length lt dim post dat 1 EO a on Islength 4 Xri postdat ri CSM Region 1 ebe pest ddtla y Algora chm Carec 4 temprow lt post dat i temp matrix rbind temprow langcoord caret langcoord caret lt temp matrix END OF FOR LOOP 4 117 itt HH HH HH HHH HH HH OH EE HOH LLL SPM2 LANGUAGE SITE COORD itt HH Ht Ht HH HHH HH HH OH HO EE H langeoord som lt post dat EE Lemp Matrix lt Jlangcoord spm length lt dim post dat 1 for j in I length 1 TL post dat j CSM Regron 1 6 amp post dat j Algorrthm SPM2 3 4 temprow lt post dat j temp matrix rbind temprow langcoord spm langooord spm lt Lemp matrrx END OF FOR LOOP f TETTE ae A He a dd E AE AE dg dd dg dd dd dd FE TE dd E E EET RETI CREATE LIST OF LANGUAGE SITES SEPARATED
102. malization methods including 1 fifth order polynomial warp 11 discrete cosine basis functions and 111 a movement model based on full multi grid approaches support Hellier s findings When Crivello et al used the four methods to normalize 20 subjects MRIs and PET volumes to the Human Brain Atlas HBA they found the full multi grid approach due to the large number of DOF provided enhanced alignment accuracy as compared to the other three methods The fifth order polynomial warp and discrete cosine basis function approaches exhibited similar performances for both gray and white matter tissues and the affine approach had the lowest registration accuracy Crivello er al 2002 Many authors refer to affine transformations as linear This is not strictly true as a linear map is a special map L that satisfies L ax Bx aL x BL x where Vx x a T and x any point in a mapping The translational part of an affine transformation violates this Thus an affine map is more correctly defined as the composition of linear transformations with translations Hill et al 2000 Grachev s anatomically based assessment of the Talairach stereotaxic transformation Talairach and Tourneaux 1988 a piece wise affine algorithm and a fifth order polynomial transformation Woods et al 1998 revealed that both methods located about 70 of anatomical landmarks with an error of 3 mm or less For landmark accuracy less than or equal to 1 mm the
103. ment vent COIT X XXXXX tOpO 9 1 1 4 Cut Topo PXXX EXXXXX SX 111 L full segment vent corrX CUT xxxxx topo 9 1 1 5 Fiducial Coord PXXX EXXXXX SX III L full segment vent corrX fiducial magctr xxxxx coord 9 1 1 6 Spherical Coord PXXX EXXXXX SX III L full segment vent corrX SPHERE Std xxxxx coord 9 1 1 7 Flat Coord PXXX EXXXXX SX 111 L full segment vent corrX FLAT Cartesian xxxxx coord 9 1 2 Atlas tab 9 1 2 1 Spec COLIN L LANDMARKS REG with INDIVIDUAL CORE6 Human colin L REGISTER with INDIVIDUAL COREO xxxxx spec 9 1 2 2 Border Human colin L LANDMARKS REG with INDIVIDUAL COREOG xxxxx borderpro 9 1 2 3 Closed Topo Human colin Cerebral L CLOSED xxxxx topo 9 1 2 4 Cut Topo Human colin Cerebral L CUTS xxxxx topo 9 1 2 5 Fiducial Coord Human colin Cerebral L FIDUCIAL SPM2 03 12 xxxxx coord 9 1 2 6 Spherical Coord Human colin Cerebral L SPHERE STD xxxxx coord 9 1 2 7 Flat Coord Human colin Cerebral L FLAT CartSTD xxxxx coord 96 9 Caret normalization of individual surface to atlas continued 9 1 3 Spherical Parameters tab 9 1 3 1 Read Params from Deformation Map File J COLIN L LANDMARKS REG with INDIVIDUAL _CORE6 TEMPLATE REG with POP AVG 4K NoFid deform map 9 2 Click OK to launch the deformation This process takes 15 30 minutes depending on the number of nodes and processing speed of your computer 9 3 If you get a dialog reporting 12 or more crossovers then something probably has gone wrong Che
104. ne T as the spatial transformation mapping from source image to target image We define T as the transformation that maps both position and intensity The first category of spatial normalization methodology employs feature based matching techniques Normalization algorithms that make use of geometrical features in images such as points lines and or surfaces determine the mapping of T positional normalization transformation by identifying features such as sets of image points that correspond to the same physical entity visible in both images and calculating T for these features Such algorithms iteratively determine T and then infer T intensity normalization transformation from T when the algorithm has converged For the purposes of this paper we will refer to methods that use this type of algorithm as surface based The second category employs volumetric transformations involving intensity values These normalization algorithms iteratively determine the image transformation 7 that optimizes a voxel similarity measure We will refer to such methods as volume based Hill et al 2000 Both surface based and volume based normalization methods may employ a rigid body transformation in other words there are six degrees of freedom DOF in the transformation three translations and three rotations The key characteristic of a rigid body transformation is that all distances are preserved Rigid body transformations ignore tissu
105. ne method or common to both for a given site mapping This analysis includes 105 or 84 of the 125 total errors Of the 38 common errors 28 20 7 of all errors site mappings for both methods were assigned the same error type table 9 The remaining ten common errors were assigned different error types depending on the method If the 28 same error type mappings had been correctly mapped 12 of the sites would have been located on the superior temporal gyrus STG four on the angular gyrus AnG and four on the precentral gyrus PrG accounting for more than 71 of the same error type mappings Table 9 Same Error Type Mappings CSM sites mapped incorrectly by both methods and assigned the same error types Language sites are in bold with green background ALD SPM2 MSTG no MMTG 1 3 no MMTG AMTG MMTG l MMTG MSTG 1 3 no MMTG 1 3 no MMTG VPrG 3 no VPoG 3 no VPoG MITG 0 5 Zz sulcus 0 5 2 sulcus MSTG l 3 no MMTG l 3 no MMTG AnG 0 25 l PMT 0 25 I PMTG P61 26 MMTG 0 25 I PMT G 0 25 l PMTG P62 28 MSTG 3 no MMTG 3 no MMTG VPrG 3 no VPoG 3 no VPoG OpIFG 0 25 l VPrG 0 25 i VPrG ASTG 1 3 no AMTG 1 3 no AMTG AMT 0 25 l MMTG 0 25 l MMTG MMTG 0 25 l AMTG 0 25 l AMTG 53 The same error type mappings if scored using the protocol outlined in Section 3 step 6 represent a total deduction of 18 5 points If these deductions were credited to the actual scores
106. ngth tempRowNum if length RowNum vec gt 1 tempRowNum c RowNum vec tempRowNum RowNum vec lt tempRowNum totalLength lt totalLength tempLength END OF FOR LOOP itt HH HH HH HH HH HH oH HH HOE HOO EE OEE EES EEE EE EEE du PRENORM DISTANCES BETWEEN LANGUAGE SITES W IN EACH BRAIN itt HHH Ht HH HHH HH HO HEH HOE EH HOE EE SEE ESE EEE d ESE EEE du prenorm lang list lapply lang prenorm list ftunction cda 1 dust lt prenorm matrix cbind tod X codoY coeA dist dist prenorm matrix method euclidean j tt Ht tH HHH EH HE EH HEH HEE EH EE TE EE HH EEE EE HH EEE E EE EE EH HH EE H CARET DISTANCES BETWEEN LANGUAGE SITES w in EACH BRAIN ttt HH HE EE EE HEH HH HH EEE EE EH HH EE HE EH HE EEE E E E EE E E E EE HE lang caret lrst lt split langecord caret hangcoord caret SBrain 1D SLR Carel dist lt apply lang Galete Listy function cd 4 domb careLmnabtrix lt obind ogdsX COd5Y 0094 dist dist caret matrix method euclidean tt tt tH HH EH HE HE HH EH HEE EH EE EH HE EEE EH HE EE E E EE EE EH HE EE SPM2 DISTANCES BETWEEN LANGUAGE SITES W in EACH BRAIN tt tt HH HH EH HE HH HE EH HEE EH EE EH HH EEE EE EH HH EEE EE EE HH EE lang spm list lt split langcoord spm Langcoord spmoBrain rcb DEENEN som last lt lappry lang spm l1865 function cd di stsomumetri1x x Corned edsx codoY CdS dist dist spm matrix method euclidean WW itt Ht tH HHH
107. o Standard Orientation dialog and switch the main caret window to the Fiducial surface Select Toolbar M for medial view 5 34 Surface Measurements Generate Curvature to update the surface shape Select Folding for the folding column and Gaussian for the gaussian column Notice the medial wall looks smooth 5 35 File Save Data File Surface Shape file Pxxx Exxxxx Sx lll L fullsegment vent corrx xxxxx surface shape and overwrite the existing file 5 36 Exit Caret 85 6 Translate cropped volume to magnet coordinate space 6 1 The current fiducial surface reflects the cropped volume left hemisphere only Pxxx Exxxxx Sx_111 L full sMRI mnc Thus it must be translated by Xmin Ymin Zmin to reflect the grid of the uncropped volume Pxxx Exxxxx Sx_111 LR full sMRI mnc left and right hemispheres 6 2 Get XYZ min from the params file grep min Pxxx Exxxxx Sx_111 L full sMRI params Look for these lines in the output Xmin 50 Ymin 23 Zmin 9 6 3 Get the magnet center from the uncropped volume mincinfo Pxxx Exxxxx Sx_111 LR full sMRI mnc dimension name length step start zspace 229 I 107 6 yspace 229 I 114 5 xspace 229 I 117 7 6 4 Translate the fiducial surface Xmin Ymin Zmin then 117 7 114 5 107 6 like so 6 4 1 6 4 2 6 4 3 6 4 4 6 4 5 6 4 6 Change to the Pxxx SURFACES and open caret5 Select the REG with Colin Core6 spec file Accept default spec file selections Make sure main w
108. on is Pxxx Exxxxx Sx_111 L full segment vent corrX mnc then don t delete that volume but delete these Pxxx Exxxxx Sx lll L full segment vent corrl mnc Pxxx Exxxxx Sx_111 L full segment vent corr2 mnc We save these too although they re rarely needed once you get to this point Pxxx Exxxxx Sx 1ll L full RadialPositionMap mnc Pxxx Exxxxx Sx_111 L full segment mnec Pxxx Exxxxx Sx lll L full segment vent corr mnc Pxxx Exxxxx Sx lll L full segment vent mnc 8 Drawings borders in Caret for surface normalization Note Also see the Spherical Registration section in Caret5 User s Manual and Tutorial Version 5 1 April 9 2004 page 61 Some differences e colin SPM2 fiducial is used in lieu of 711 2B version e Core6 landmark set is used e adds superior temporal gyrus e border extents avoid sulci margins e uses parameters in deformation map provided in Appendix G 88 8 Drawings borders in Caret for surface normalization continued 8 1 You will need two caret sessions open side by side One with the PX XX and one with the target atlas in this case the target atlas is colin 8 2 Each session will have the flat map in the main window with the inflated surface in window 2 Make the windows as big as your screen will allow 8 3 First you ll click green ID nodes on the inflated surface to get an idea of where to start and stop drawing each border We don t draw to the end of each sulcus because near the margins the co
109. on originally developed the software and associated theory for routine statistical analysis of functional neuroimaging data SPMc assic was the first version of the software suite released in 1991 with the intent of promoting collaboration and a common analysis scheme across 2 laboratories SPM has had five major revision releases since 1991 In this study we consider the most recent release SPM2 released in 2003 Spatial normalization using SPM2 is achieved by registering the individual MR images to the same target image by minimizing the residual sum of squared differences between them The first step in spatially normalizing each image involves matching the image by estimating the optimum 12 parameter affine transformation Ashburner et al 1997 A Bayesian framework is used whereby the maximum a posteriori estimate of the spatial transformation 1s made using prior knowledge of the normal variability of brain size This step has been made more robust in SPM2 Affine registering image A to match image B should now produce a result that is much closer to the inverse of the affine transformation that matches image B to image A A regularization a procedure for increasing stability of the affine transformation has also changed The penalty for unlikely warps is now based on the matrix log of the affine transform matrix being multivariate and normal The second step accounts for global nonlinear shape differences which are modeled by a l
110. ost residual errors will require some type of surface patching The tools for patching include the following 4 11 1 4 11 2 4 11 3 Toggle Voxels Press the Toggle Voxels button to switch the mouse buttons to the voxel editing mode While in this mode the left mouse button still positions the cursor but the middle button makes the voxel white and the right button makes it black If multiple voxels need editing the process can be expedited using the arrow keys to positions the cursor with one hand while modifying voxels with the other hand on the mouse Note that the voxel affected is determine by the position of the cross hairs and NOT the cursor Zoom and pan are disabled until you press the Resume Normal Mouse Mode button so be sure to exit the Toggle Voxels mode before resuming other operations Also you will want to check the box for apply to current slice only so that your changes are limited to one slice at a time versus all slices at once Masking The mask function allows you to toggle more than one voxel on and off at once You can select the mask dimensions and place the mask using the cross hairs and select dilate or erode depending on what 1s required for a given error correction Flood Filling This option is used for removing disconnected regions If a finger has been disconnected by deleting voxels that link it to the main segmentation press the Flood Fill Volume 2 button The largest segmented object will rema
111. parison of three independent mappings of two subject s CSM sites n 41 sites that any given mapping could vary by an average of 7 5 mm still within the 1 cm margin of error for site location mapped during surgery 16 1 6 Hypothesis In this case study we expected that reducing variation would better reveal functional patterns of language production that exist in CSM language sites which have been identified as statistically significant for naming errors during neurosurgery We believed that the problem of analyzing CSM functional data across subjects can be solved using computer aided spatial normalization As aresult we asked this key question What is the best spatial normalization method for registering two or more brains such that the observed variation in the functional areas after registration as measured by cortical stimulation mapping CSM is the smallest Because the data we used in our case study was collected on the cortical surface we expected that a surface based normalization method which relies on selected cortical surface landmarks would result in less variation between CSM language sites as compared to a volume based normalization method For the same reason we expected that the surface based normalization method would result in more accurate anatomical localization of CSM site locations 17 Section 2 Survey of Evaluation Tools Developing and implementing an evaluation technique for spatial normalization
112. rative photo Ojemann et al 1989 13 Once the craniotomy is complete local anesthesia 1s applied to the dura and scalp and the patient is brought to an awakened state in preparation for stimulation mapping typically three to four hours after the operation has begun The awakened patient is shown slides of line drawings of familiar objects like planes boats trees etc The slides are projected at four second intervals with the patient trained to name each one as it appears This is an easy task and there are frequently no naming errors on slides presented in absence of stimulation Out of the 117 subjects included in the Ojemann study the highest error rate without stimulation was 22 While the patient names the slides sites identified by numbered tags are successively stimulated with the current applied as the slide appears and continuing until the appearance of the next slide At least one slide without stimulation separates each stimulation and no site is stimulated twice in succession Usually several slides intervene between each stimulation and all sites are stimulated once before any site is stimulated a second time Three samples of stimulation effect are usually obtained Intraoperative manual scoring of errors and their relations to stimulation provides immediate feedback to the neurosurgeon Ojemann et al 1989 If the stimulation of a site results in a naming error at least two of the three times the site is determined to be ess
113. re these files are NOT selected CUTtopo file caret P117 E10043 S4 111 L full segment vent corr3 FLAT Cartesian 71785 topo FIDUCIALcoord file caret P117 E10043 S4 111 L full segment vent corr3 fiducial magctr 71785 coord FLATcoord file caret P117 E10043 S4 111 L full segment vent corr3 FLAT Cartesian 71785 coord surface shape caret P117 E10043 S4 111 L full segment vent corr3 71785 Surface shape 12 4 Load the selected files 12 5 Ifthe foci file was in the spec file when registration was performed then it was deformed during registration Otherwise apply the deformation map as follows 12 5 1 Surface Deformation Apply Deformation Map caret P117 EI0043 S4 I11 L REG with Colin Core6 2004 09 14 11 03 71785 deform map 12 5 2 File Type Foci 12 5 3 Data file Pxxx SURFACES Pxxx_CSM foci 12 5 4 Apply 12 5 5 Close the deformation dialog 12 6 Select File Open Data File type Foci caret Pxxx CSM foci 12 7 Replace any existing foci 12 8 If CSM focicolor wasn t in the spec file File Open Data File Foci Color CSM focicolor 12 9 Toolbar L to switch to lateral view 104 12 Caret normalization of CSM coordinates to atlas coordinates continued 12 10 Toolbar D C and toggle on Foci 12 11 Layers Foci Project Fiducial Foci Hemisphere only keep offset from surface 12 12 File Save Data File Foci Projection caret Pos CSM fociproj 12 13 D C On the Shape drop down menu make sure Depth is the selected column if this op
114. results and a slight degradation of the SPM2 normalization results Caret reduced the spread between sites by 5 1 mm more than SPM2 in 2D space In 3d space Caret reduced the spread by 1 9 mm more than SPM2 Using the jackknife estimate of variance method we found that this difference remained statistically insignificant However the difference in the means show that a Caret mapping will on average be better than the SPM2 mapping by more than 5 mm in 2D space and almost 2 mm in 3D space Also the confidence interval revealed that a Caret mapping could be as much as 13 mm better than a SPM2 mapping in 2D space and more than 6 mm better than a SPM2 mapping in 3D space A power t test calculation was used to estimate the number of subjects required to achieve a statistical significance of p lt 05 and 80 power We found that for 2D analysis we would need at least 55 subjects For 3D analysis 120 or more subjects would be required 58 Section 5 Discussion Anatomical Variation between Source and Target The common mapping errors support what visual inspection of the structural surfaces of both the source and target hemispheres revealed locations of structural vagaries in both the colin27 and in our subjects average surface reconstruction were where mapping error rates were 50 or greater figure 31 The colin27 atlas structural regions were observed by a neuroanatomist to be atypical in the ventral portion of the precentral gyrus VPrG
115. rface reconstruction 3 create flat map 4 assign coordinates function and cortical parcellation to each CSM site 5 apply spatial normalization to anatomical and functional data 6 evaluate methods using spread reduction and anatomical localization measures 29 Step 1 Select MRI volumes Figure 7 Visual Brain Mapper screen shot Upper left is a neurosurgery photo of the left temporal lobe with sterilized labels identifying various cortical sites To the right are coronal and axial slices from the patient s MRI taken prior to surgery At lower left is the lateral left hemisphere view of a 3D brain model including arteries and veins created from the patient s MRI venogram and arteriogram We selected 11 MR images from a University of Washington Structural Informatics Group database of over 90 patients CSM database We screened the database images for left hemisphere surgery quality and lesions The first level of screening eliminated patients whose surgery was conducted on the right hemisphere By including only left hemispheres in this case study we limit our scope to focus on one major structural element of the brain While left hemisphere surgeries are more common than right hemisphere surgeries future work would need to include analysis of the right hemisphere as well as the left Image quality was the next level of screening We determined quality by uniformity of voxel intensity values gray white contrast within th
116. roimages Computers and Biomedical Research 1996 29 162 173 Crivello F Schormann T Tzourio Mazoyer N et al Comparison of spatial normalization procedures and their impact on functional maps Hum Brain Mapping 2002 Aug 16 4 228 50 Crum WR Griffin LD Hill DLG et al Zen and the art of medical image registration correspondence homology and quality NeuroImage 2003 July 20 1425 1437 D Agostino E Maes F Vandermeulen D Suetens P A viscous fluid model for multimodal nonrigid image registration using mutual information Lecture Notes in Computer Science Proc Medical Image Computing and Computer Assisted Intervention 2002 Sep 2489 541 548 Davatzikos C Bryan RN Using a deformable surface model IEEE Transactions on Medical Imaging 1996 Dec 15 6 789 795 Davatzikos C Li HH Herskovits E et al Accuracy and sensitivity of detection of activation foci in the brain via Statistical Parametric Mapping a study using a PET simulator NeuroImage 2001 13 176 184 Davatzikos C Bryan RN Morphometric analysis of cortical sulci using parametric ribbons a study of the central sulcus J Comput Assist Tomogr 2002 26 2 298 307 Desai R Liebenthal E Possing ET Binder JR Volumetric vs Surface based Intersubject Alignment for Localization of Auditory Cortex Activation CNS Poster 2004 Efron B and Tibshirani RJ Chapter 11 An Introduction to the Bootstrap Chapman amp Hall 1993 Fischl B Sereno M Dale
117. rrespondence becomes less clear between the individual s and colin s folding patterns For example colin has a branch at the dorsal tip of his central sulcus whereas most subjects don t By starting at the point where this branch merges and beginning a similar distance from the medial wall in the individual we can be confident that these landmarks correspond to one another 8 4 For the atlas session open Caret5 in COLIN L LANDMARKS REG with INDIVIDUAL CORE6 directory 8 5 Select caret Pxxx Exxxxx Sx 111 L REG with Colin Core6 xxxxx spec 8 5 1 Select Geom atlas flat and inflated surfaces 8 5 2 Select Border border color and Core6 border projection files 8 5 3 Select D C surface shape Mean Curvature Folding 8 6 For the PXXX session open Caret5 in PXXX SURFACES directory 8 7 Select Geom INFLATED SPHERE Std and FLAT Cartesian surfaces 8 8 Select Border LANDMARKS FromFlattening borderproj file 8 9 Select Toolbar Spec 8 10 Select LANDMARKS FromFlattening borderproj REPLACE existing borders File Open Data File Border color file and navigate up and over to the colin directory 8 11 Select ForSPHERICAL REGISTRATION Human Class3 bordercolor and copy the file to the existing directory 8 12 Select Layers borders project borders nearest tile Border points on the origin may disappear 89 8 Drawings borders in Caret for surface normalization continued 8 13 In both sessions do the following 8 13 1 8 13
118. rs on the flat map s folding pattern even if your ID marks miss the fundus The SF STSant landmark is a gyrus so aim for the whitline The green ID nodes show you where to start and stop drawing Draw borders as follows 8 20 1 8 20 2 8 20 3 8 20 4 8 20 5 8 20 6 8 20 7 Layers Border Draw Border Name LANDMARK CentralSulcus Resampling 4 0 Apply Starting at the ventral tip of the central sulcus near the origin scale bar near the center of the surface draw the central sulcus by dragging the mouse with the left mouse button depressed When you get to the dorsal extent delimited by your green ID node shift click the left mouse button to terminate the border On the Draw Border dialog select LANDMARK SylvianFissure from the Name menu and click Apply Starting at the dorsal tip of the Sylvian fissure just across from the origin scale bar near the center of the surface draw the Sylvian fissure by dragging the mouse with the left mouse button depressed When you get to the ventral extent delimited by your green ID node shift click the left mouse button to terminate the border On the Draw Border dialog select LANDMARK SF STSant from the Name menu and click Apply Starting at the dorsal tip of the superior temporal gyrus closer to the center of the surface trace along the white line until you reach the end point then shift click 8 20 8 8 20 9 If you re not happy with a border select Layers
119. s are defined midbrain pons and medulla In both cerebrum and cerebellum brain areas lying deep in traditionally defined lobes are termed sub lobar Level Three gyrus divides each lobe into gyri or gyral equivalents Nuclear groups such as thalamus or striatum are gyral equivalents Level Four of the hierarchy is tissue type Each gyrus or gyral equivalent is segmented into grey matter white matter and CSF Level Five of the hierarchy is cell population Cerebral cortex is labeled by Brodmann area Nuclear groups are labeled by subnuclei Cytoarchitectonic labels for cerebellar cortex and tract labels for white matter are being developed but are not yet available The Talairach Daemon s labels are stored as a volume array 1 mm isometric voxels spanning the extent of the brain in the Talairach 1988 atlas This corresponds to approximately 500 000 voxels Each voxel in this array contains a pointer to voxel specific brain information This information is called a relation record and is managed as a linked list A relation record can store any information that is recorded using Talairach coordinates To eliminate the need for storing 22 duplicate information in relation records each record contains pointers to the information rather than the information This scheme offers the potential for extremely high speed access to information within the relation records Lancaster et al 1997 MNI305 and ICBM152 The Montreal Neurologic
120. s associated with them All localization endeavors were given a confidence rating on a scale from 1 to 5 where 1 is not at all confident and 5 is very confident The ratings were determined by amount and quality of images and descriptions Using the blood vessels and anatomical structure in the rendering as landmarks the expert drags and drops the number that corresponds to the number in the photo onto the rendered image Once the site has been dropped a pick operation is performed in order to determine the closest surface facet to the site The site is assigned a 3D coordinate in MR magnet space in which the center of the MR magnet is the origin This data is stored in a cortical stimulation mapping CSM database 15 IO Superior 21 7 tk Anterior 6 9553 is NOT lanquage essential Camera Controls Light Ca IE Number Sizes MRI Window Width Leve Isosur ace experiments Match Eyepoints SAVE MAP Corlev color Artery color L fe e E Vein color Figure 4 Skandha4 GUI used in the visual comparison approach According to repeatability studies any given mapping will typically fall within a distance of 5 1 mm of the true site location as measured by the mean of all the mappings 6 mappings included in the study Since the language site locations mapped during surgery are accurate to 1 cm the accuracy achieved using the visual comparison approach was deemed satisfactory Modayur et al 1997 We found in a com
121. s the medial wall cut Five other cuts include calcarine cingulate frontal lateral and temporal drawn in blue Once the cut lines were set the automatic flattening took place Flattening took 30 60 minutes on a Dell Dimension 450 with dual processors running Debian Linux Figure 10 Relationship of 3D surface reconstruction to 2D flat map 34 Step 4 Assignment of Coordinates Function and Cortical Parcellation to Sites In Section 1 5 we described how cortical stimulation data was collected during neurosurgery and mapped to a coordinate system using the visual comparison approach Additionally the neuroanatomist expert assigned an anatomical location to each site based on a cortical parcellation system CPS designed as a scheme for examining the neural substrate through intelligent computer querying of the CSM database Corina et al 2005 This system divides the lateral surface of the cortex into 37 subdivisions labeled using the Foundational Model of Anatomy FMA expansion of NeuroNames terminology and is shown in figure 11 Anatomical AKA sulcal boundarv Subjective boundary Figure 11 Cortical Parcellation System for lateral cortical surface The data retrieved from the CSM database included the 3D coordinates and CPS anatomical localization for each of the 198 sites recorded for the 11 subjects The coordinate file was then input into both the surface based and volume based methods and transformed accordingly In
122. sing steps 79 2 Volume Preparation continued 2 4 Select Anterior Commissure 2 5 2 4 1 2 4 2 2 4 3 2 4 4 2 4 5 The view will automatically switch to coronal view Move the parasagittal red line cursor to the midline Switch to the parasagittal view Center the blue and green cross hairs on the anterior commissure see reference photo in SureFit Switch back to coronal view and adjust parasagittal cursor to intersect the midline precisely at this coronal view Press Set Anterior Commissure button Select Define VOI and Identify Cut Faces Ziad 2 93 42 B 2 5 4 ZO 2 5 6 Zot 29 6 Note The SureFit segmentation algorithm currently works only on hemispheres and portions thereof it does not work on entire brain volumes You must crop to at most a left or right hemisphere before proceeding to segmentation Typically we will crop the eft hemisphere Select horizontal panel and scroll to the slice level where the hemisphere is widest and longest Adjust the min and max X slider bars to choose the medio lateral extent of the volume to be segmented For the X axis cropping several mm beyond midline into the opposite hemisphere is a good idea and prevents inadvertently clipping bits of VOI Adjust the min and max Y slider bars to choose the anterior posterior extent of the volume to be segmented Select the Crop button to apply the newly defined extent to the X and Y axes Scroll through the
123. t surgery on the left dominant hemisphere In order to insure that cortical surface sites without evoked naming errors could be resected with a low risk of postoperative language deficit stimulation mapping must indicate both where language function is located and where it is not The extent of the craniotomy is determined in part by this consideration covering both the areas of the proposed resection and also the likely language locations Ojemann et al 1989 The patient is put under general anesthesia for the craniotomy Prior to mapping rolandic cortex is identified by stimulation and the threshold for after discharge in the electrocorticogram ECoG is established for the area of association cortex to be sampled with language mapping Language mapping is conducted with the largest current that does not evoke after discharges Typically this current is in the 1 5 to 10 mA range measured between peaks of biphasic square wave pulses with a total duration of 2 5 msec 1 25 msec for each phase This current is delivered from a constant current stimulator in four second trains at 60 Hz across 1 mm bipolar electrodes separated by 5 mm Sites for stimulation mapping are randomly selected to cover the exposed cortical surface including areas where language function 1s likely located as well as proposed resection Typically there are 10 20 stimulation sites per subject that are identified with sterile numbered tags and recorded with a digital intraope
124. tation III Generate surface Initial surface reconstruction IV Correct 5 errors Corrected cortical segmentation V Generate fiducial surface d Fiducial surface reconstruction VI MapfMRI data 5 5 optional l Functional activation maps Figure 6 Five steps for creating a surface reconstruction using SureFit We completed the first five steps for each subject s left hemisphere Instead of mapping MRI data as the sixth step we mapped CSM data Other tools considered include Freesurfer Dale et al 1999 mrGray 2 0 Teo et al 1998 and BrainVoyager Goebel et al 1997 We selected SureFit primarily due to our past experience with the tool and the desire to collaborate with the Van Essen Lab 20 2 2 Surface flattening tools The complex geometry of the human brain contains many folds and fissures making it impossible to view the entire surface at once Since most of the cortical activity occurs on these folds it 1s desirable to view the entire surface of the brain in a single view This can be achieved using flat maps of the cortical surface which are essentially unwrapped cortical surfaces in a 2D plane Van Essen et al 2001 Cortical flat maps also make it easier to see the depths and complete shape of the sulci Algorithms for creating flat maps do require cutting compression and stretching of the surface causing some distortion All cortical flattening methods aim to minimize geometri
125. te matter the inner gray white boundary and the outer pial boundary as substrates for the segmentation process using Gaussian intensity transformations Van Essen et al 2001 This generation requires a complex set of filtering operations intensity transformations and other volumetric operations applied to the image intensity data All filtering operations are applied to the 3D image volume Inner and outer boundary maps are particularly important because they are combined to form a position map along the radial axis which runs from the inner to the outer boundary The result is a position map along the radial axis that is thresholded The thresholding generates an initial cortical segmentation with a boundary running approximately midway through the cortical sheet The initial segmentation is used as the substrate for generating an explicit surface reconstruction Lorenson 1987 SureFit currently involves five major processing stages for segmentation as shown in figure 6 19 SureF it Cortical Surface Reconstruction Process Volumes amp Surfaces Raw image intensity Orient Volume 1 Define Volume erer Ae SH Of Interest VOD d I PS EE optional Oriented cropped intensity volume Set Parameters Generate probabilistic volumes Composite inner Composite outer I II i boundary boundary I KI E Radial position map Segment volume Initial cortical segmen
126. tex depending on the subject 1 4 Survey of Spatial Normalization Methods Computer aided spatial normalization is a widely used solution for relating the anatomy and functionality of multiple brains in neuroscience and is a critical step in quantitative analysis of the human brain cortex It is not practical nor desirable to completely normalize the function and structure of one brain to another Rather the goal of most researchers is to bring an individual s functional and structural data into a common visualization substrate with a set of common coordinates Having registered cortical structures one can perform group or individual analyses of structure and function to assess normal group differences in terms of age gender genetic background handedness etc Ashburner er al 2003 Davatzikos and Bryan 2002 Mangin et al 2003 Thompson et al 2001b We can also better define disease specific signatures and detect individual cortical atrophy based on computational anatomy methods May et al 1999 Thompson et al 2001a Toga and Thompson 2003 Other applications for spatial normalization methods include automatic cortical structural labeling and visualization Le Goualher et al 1999 functional brain mapping Toga and Mazziotta 2000 and neurosurgical planning Kikinis et al 1991 Given the wide array of intersubject registration applications many image analysis methodologies have been developed to address this need We defi
127. tically insignificant difference in overall accuracy between the surface based method resulting in 63 errors and the volume based method s 62 errors when mapping a total of 198 sites Qualitative analysis of the error types provides more insight into some common and unique problems of spatial normalization in this case study Most notably of the 125 total errors 38 sites 60 of total errors were incorrectly mapped by both methods Also a paired t test showed a statistically significant difference in the type 2 errors mapped by both methods While SPM2 normalization resulted in only one type 2 error Caret normalization resulted in 18 such errors 2D Language Site Spread 2D Spread Reduction e D E E E E st f EN Expected CARET SPM 2 CARET SPM 2 Figure 21 2D analysis of mean distance between 21 language sites across 11 subjects 3D Language Site Spread 3D Spread Reduction Ui D E E E E p Expected Caret SPM 2 CARET SPM2 Figure 22 3D analysis of mean distance between 21 language sites across 11 subjects 47 Table 7 Caret Error Type Summary ID Type 1 Error Type2 Error Type3 Error Total Error 54 0 4 3 7 S P 2 4 8 58 1 l 2 4 60 1 2 5 8 61 4 2 l n 62 l l 2 4 63 2 1 7 10 117 0 p 1 3 164 0 1 2 3 170 4 2 0 6 176 3 0 0 3 Totals 18 18 27 63 Table 8 SPM2 Error Type Summary ID Type 1 Error Type2 Error Typ
128. time or in small clusters using dilation and erosion steps within small masked regions Error correction was completed when no visible handles remained on the cortical surface reconstruction The quality of the SureFit generated cortical segmentations was evaluated by visual inspection of segmentation boundaries and of surface contours overlaid on the anatomical volume This assessment suggested that surfaces are generally accurate to within about 1 mm of their desired trajectory Van Essen 2005 32 Step 3 Create Flat Map To aid in method evaluation we created cortical flat maps using Caret as outlined in Section 2 2 The SureFit specification file for the individual surface reconstruction was loaded into Caret The Flatten Surface functionality was selected Then the six default cuts outlined on the medial surface of the left hemisphere were inspected The calcarine cut and medial wall cut were always redrawn to match the specific structure of the individual surface The remaining cuts cingulate cut frontal cut Sylvian cut temporal cut were redrawn as needed using the Draw Border functionality These cut lines were used to determine where the inflated surface was split in order to achieve a cortical flat map Figure 9 shows the template cut lines on the surface reconstruction and figure 10 shows how the surface reconstruction and flat map correlate 33 Figure 9 Template cuts for flattening The red dashed line trace
129. tion isn t available make sure you have Human colin Cerebral L xxxxx surface_shape loaded not the deformed Pxxx surface shape file 12 14 D C Foci menu Draw Foci as Spheres adjust foci size as desired 12 15 Switch to flat view 12 16 File Save Window as Image caret_Pxxx_flat jpg 12 17 Switch to inflated fiducial views and save those captures if desired 12 18 Save flat foci coords as follows 12 18 1 File Save Data File Foci File 12 18 2 Foci Associated with Surface Type Flat 12 18 3 Filename caret_Pxxx_CSM_flat foci 12 19 caret Pxxx CSM foci and caret Pxxx CSM flat foci files will be input into the PostNorm csv and Flat Postnorm csv files for calcuation of spread reduction see Appendix F 105 13 SPM2 normalization of CSM coordinates to atlas 13 1 Create a coordinate file stripped of header and node numbers in Pxxx directory Name file Pxxx coord stripped See example of stripped coordinate file in Appendix C 13 2 Move the norm coord m script Appendix D to the individual directory Pxxx 13 3 Start matlab in Pxxx directory 13 4 Run norm coord Pxxx coord stripped which will call SPM2 13 5 Selectiy Pxxx Exxxxx Sx lll img file 13 6 Output should be spm Pxxx coord stripped file in Pxxx directory It will be a coordinate file stripped of any headers and comments 13 7 Once you have the SPM2 normalized CSM coordinates file you can use the merge spm foci sh script Appendix E at a Linux command line to gener
130. tion of anatomical and functional variation increases the distance between language sites across patients It follows then 1f anatomical variation is reduced the distance between language sites across brains will be reduced Therefore we expect that the distance between language sites across patients what we will refer to as spread will get smaller after spatial normalization The optimal method will maximize spread reduction This hypothesis assumes that the volume and surface areas of the source and target hemispheres are the same As we can see in Table 5 the mean surface and volume areas of the 11 subjects are less than the target s surface and volume areas To accommodate for this difference we calculated the ratio of the mean subject volume and surface area to the corresponding colin27 volume and surface area 42 Table 5 Surface and volume areas of the 11 subjects an atlas left hemispheres Left Hem Gender Volume mm 3 M 559616 512958 606962 466446 510960 448175 465153 498930 554765 423512 409783 496115 714773 M M M F F M F M F F M Given the difference between the volume and surface areas of our source and target hemispheres we calculated an expected change in post normalization distances We used the following values to estimate the expected post normalization distance EPoD PrD average pre normalization AC PC aligned distance between 21 language sites CSA colin27 surface area C
131. tributable to noise large blood vessels or regional inhomogeneities in the structural MRI volume or a combination of these Errors were localized by inflating the initial surface reconstruction to a highly smoothed ellipsoidal shape and using the orientation of surface normals to identify regions called crossovers where the surface is folded over itself Clusters of surface nodes associated with crossovers were mapped from the surface reconstruction into corresponding voxel clusters in the volume The automated error correction process tested for different types of handles in the vicinity of each location determined to have an error The localized patches used for these tests conformed to the shape of temporary segmentations that are based on different threshold levels for the radial position map If the trial patch reduced the number of topological handles in the segmentation as determined by an Euler count applied to the volume LeeT C et al 1994 it was accepted as a permanent correction and the process moved on to the next error patch Van Essen et al 2001 The automated error correction process sometimes failed especially for handles that were notably large or irregular Such errors were corrected using interactive editing For each handle that remained after automatic error correction the analyst used the object limits and 3D viewer to identify the vicinity of each remaining handle Voxels were then added or removed one at a
132. vent corr mnc 4 3 Select SureFit run SureFit Ineractive Error Correction Update Handle Count to determine the number of handles in the segmented volume 4 4 Select Surface Operations Read Surface 1 PXXX SURFACES lt patientIDs gt _ Sx 111 L full segment vent corr fiducial lt nodenum gt vtk 4 5 Read Surface 2 PXXX SURFACES lt patientIDs gt Sx 111 L full segment vent corr inflated lt nodenum gt vtk 4 6 Select Surface Operations Paint Surface and open PXXX SURFACES lt patientIDs gt Sx 111 L full segment vent corr errors lt nodenum gt RGBpaint 4 7 Select SureFit run SureFit Interactive Error Correction tab Click on Locate Objects button You will be prompted for a minc file If Vol 2 was run through error correction then accept default errors file to bring up a list of object locations in the form xmin xmax ymin ymax zmin zmax These are the limits of objects that the error checking algorithm has flagged as problem areas 4 8 In the slice window scroll to the first set of specified limits for each dimension x y z 4 9 Press the keyboard letter p and rotate fiducial and inflated surfaces to find the red dot over the red surface patch representing the error 4 10 Scroll through the volume in the slice window within the area of the error toggling between Vol 2 and Vol 1 amp 2 images to identify details and the nature of the residual error 80 4 Interactive error correction continued 4 11 M
133. ween language sites across subjects which is necessary in order to measure spread reduction after normalization Step 6 Each subject s volume has its own magnet center and in some cases the chin may be rotated up or down or slightly to the side For pre normalization coordinates we aligned the anterior commissure AC and posterior commissure PC using the AFNI software package This process resampled the volume to cubic 1 mm voxels and applied a rigid registration to align the volumes to a common origin the intersection of the superior edge and posterior margin of the AC AFNI also rotated the volume as needed so that its Y axis runs from the inferior edge of the PC to the AC origin Using AFNI the AC superior edge and posterior margin the inferior edge of the PC and two mid sagittal points were selected AFNI then computed transformation information that it stores in the volume HEAD files Then the AFNI command line utility Vecwarp was called to apply the transform to the individual coordinates resulting in pre normalization AC PC aligned coordinate files These coordinate values were used to calculate the pre normalization distances Euclidean distance between language sites across subjects measure described in step 6 37 Step 5 Spatial Normalization Figure 12 Visualization of the normalization process Surface based normalization To normalize the individual surface source to the colin27 atlas target we first need to s
134. x it i Ht Ht HH HH HE HE EEE EEE E E R E E R E E R EEE EEE EEE HHH HHH FF CREATING DATA FRAME FOR STATISTICAL ANALYSIS it it Ht Ht Ht HH HH HE EEE EEE EEE EGE EGET EE EEE EEE EHH HHH TT jackknife estimate of variance ids lt make id pairs brainlistSBrain ID uniqueids lt unique brainlistSBrain ID differences DAB vec lt caret DAB vec spm DAB vec jackknife diffs sapply uniqueids function i mean differences DAB vec idsSidl i K ids id2 i mean differences lt mean differences DAB vec se differences lt sqrt var jackknife diffs 10 Ci mean lt mean differencestc 1 96 1 96 se differences p value 2 pnorm abs mean differences se differences lower tail FALSE 116 eH Ht tH H Ht EE HH HEE HH HEE HH Ht EE HH HEE HE HH EE HE HH EE EH HE EE HH EE EH HE EH HE EE HH EE HH Ht EE HH CALCULATING AVERAGE EUCLIDIAN DISTANCE BETWEEN LANGUAGE SITES ACROSS 11 HUMAN BRAINS DATE 02 25 05 AUTHOR VERONICA SMITH HH HH Ht HH HH EE HH EE HH tH HH EE HH EE HH EE HH EE HH HEE HH EE HH EE HH HE HH EE HE HH Ht Ht EE HEHEHE HHH HHH EHH EH HEE PREP WORK EHE EHE EHE E EE EHH EH HEE CLEAN UP rm list l1s all TRUE IMPORTING DATA Presdat lt peadgcsv PreNorm acpe csv post dat lt read csv PostNorm csv pp dat lt rbind pre dat post dat SUMMARIZING DATA summary pre dat summary post dat summary pp dat LOAD 3D PLOTTING PACKAGE library scatterplot3d
135. y established by comparison of techniques reliability and computational efficiency In this evaluation study conducted as part of the University of Washington Human Brain Project we address construct validity by comparing a surface based and volume based normalization method and establishing an expected accuracy measure for the given data set We also address face validity by visual assessment of functional transformation to anatomical substrate Computational efficiency reliability and overall cost benefit analysis are also discussed 12 1 5 Cortical Stimulation Mapping and the Visual Comparison Approach Since language function has been shown to vary significantly across patients the technique of intraoperative stimulation mapping is used in order to plan for treatment of temporal tumor or intractable epilepsy at the University of Washington Devised by Penfield and Roberts stimulation mapping was based on the observation that applying a current to some cortical sites blocked ongoing object naming although no effect of stimulating these sites was reported by the quiet patient Based on this work stimulation mapping for language localization became an accepted part of resective surgical technique for epilepsy Stimulation mapping is used for localizing language function within a hemisphere after lateralization has been determined preoperatively by the intracarotid amobarbital perfusion test Our data set consists of 11 patients who underwen
136. ysis and identification of linguistic sensory stimuli named after Wernicke Wernicke 1874 Although many researchers generally accept this basic scheme there is not universal agreement on many of the details as well as whether or not Broca and Wernicke s areas are truly canonical Binder er al 1997 Ojemann er al found in their electrical stimulation mapping investigation of 117 epilepsy patients that the generally accepted model of language localization in the cortex needed revision The combination of discrete localization in individual patients and substantial individual variability between patients found in the study demonstrated that language cannot be reliably localized based on anatomic criteria alone Ojemann et al 1989 Adding to the complexity of language function is that key findings in neuroscience and cognitive science have shown that learning experiences change the physical microstructure of the brain which alter its functional organization According to Bransford New synapses junctions through which information passes from one neuron to another are added that would never have existed without learning and the wiring diagram of the brain continues to be reorganized throughout one s life Bransford et al 1999 An example of experience determining how parcels of the brain are used can be found in the brains of deaf people where some cortical areas typically used to process auditory information in hearing peopl
137. yst at the Van Essen lab currently Donna Hanlon Secondary Segmentation If there are less than 15 handles after initial segmentation then once again select SureFit Run SureFit In the Run SureFit tab use the following selections e Segmentation Scope Extract Cerebrum Segment s Fill Ventricles Yes e Generate Surface select Correct Errors and Identify Sulci Leave Keep intermediate files unselected Run SureFit button once desired options have been chosen This process will generate a cortical segmentation and associated surfaces stored in the PXXX directory including a SURFACES directory The fiducial surface will automatically appear in a separate surface viewer window VTK image when the process is complete The time to complete this process is currently approximately 2 hours on sulcus Select the L button Lateral View and save the VTK image as PXXVTK Lateral LHem jpg using the Gimp or comparable graphics software in the PXXX directory 79 4 Interactive error correction 4 1 Once the initial segmentation is run the quality of the segmentation is checked to determine if interactive error correction is required 4 2 If the volumes are not yet loaded 4 2 1 load the intensity volume as Vol 1 Volume Operations Read Volume 1 lt patientIDs gt _Sx_111 L full sMRI mnc 4 2 2 Next load the segmentation to be corrected as Vol 2 Volume Operations Read Volume 2 PXXX Segmentation lt patientIDs gt _ Sx _111 L full segment

Evaluating Spatial Normalization Methods for the Human Brain

Contents

Download Pdf Manuals

Related Search

Related Contents