Home

STUK - Julkari

1. Chakraborty DP Comparison of computer analysis of mammography phantom images CAMPI with perceived image guality of phantom targets in theACR phantom Proc SPIE 3036 160 167 1997b Chakraborty DP Sivarudrappa M and Roehrig H Computerized measurement of mammographic display image guality Proc SPIE 3659 131 141 1999a Chakraborty DP Effect of antiscatter grid and target filters in full field digital mammography Proc SPIE 3659 878 885 1999b Cunningham IA Moschandreou T Subotic V Detective quantum efficiency of fluoroscopic systems the case for a spatial temporal approach or does the ideal observer have infinite patience Proc SPIE 4320 479 488 2001 Dainty J C and Shaw R I mage Science Academic Press London 1974 Dobbins JT III Effects of undersampling on the proper interpretation of modulation transfer function noise power spectra and noise equivalent quanta of digital imaging systems Med Phys 22 171 181 1995 Dobbins J T III Ergun DL Rutz L Hinshaw DA Blume H and Clark DC DOE f of four generations of computed radiography acquisition devices Med Phys 22 1581 1593 1995 Fujita H Ueda K Morishita J Fujikawa T Ohtsuka A and Sai T Basic imaging properties of a computed radiographic system with photostimulable phosphors Med Phys 16 52 59 1989 Fukunaga K Introduction to Statistical Pattern Recognition Academic Press New York 1972 Gagne RM and Wagner RF Prewhitening matched filt
2. FluoroQuality estimates the biased SNR and SNR te of two computational observers the DC and HF suppressing observer DCsHFs and the prewhitening matched filter PWMF the ideal observer It should be noted that the residual noise in the average images and NPS data makethese esti mates high biased as discussed earlier in Chapter 2 3 and it is not advisable to use them for estimating the detectability of the signal The data may be useful however for estimating the effect of the different detection algorithms on the resulting SNR estimate The data are presented as a function of the maximum included spatial frequency in the summation f 7 Imt The SNR of the individual image frames is calculated as DCsHFs Shaun fF 0 0 0r 0 32 18 SNRocstir Fim 2 5 Yv R u v R u v AG u v f 0 0 or 0 32 where the summation is done only over frequencies f u v where 19 f Vu v S Fim and f _ is equal to 7 15 30 or not limited 31 lt u v lt 32 25 STUK A196 Theideal observer s biased SNR is calculated as AG v SNR 20 pwmr Sim 2 12 R u v R u v and again presented as a function of the maximum included frequency f In a dose analogy of the above formulae the SNR calculated as Imt of the observers is rate Show 1 f 0 0 0r 0 32 21 SNR uo 5 s Sim N ree TO Ea R moV R U VAGU V f 0 0 0r 0 32 and 22 i AG v
3. SNR is Jim T A Ya Rano u v T Rami u v 3 7 The SNR measures obtained by integration analytical de biased data As discussed by Gagne and Wagner 1998 the direct summation over the spatial frequency channels of the estimated biased SN R spectrum results to a biased SNR2 value FluoroQuality calculates also a de biased SN R esti mate by their theory as 23 SNR pai debiased Fau gt SNR seriased SSfim where 2N g f 3 AG f 24 SNR deiased z i P 4 A 2 Nag SDZ VAR PARP Ng is the bias corrected SNR spectrum and N y f is the effective number of images in either the signal or background case see the explaining text later 26 STUK A196 The error 0 of SNR voa debiasea fim IS also estimated according to the theory of Gagne and Wagner 1998 by adding the variances of each included frequency channel 25 0 SNR 4 N f 1 4 4 SNR SIR in ay Qn F NOAIN Nghe TIN The bias and variance at the spatial zero and Nyquist frequencies are actually slightly different from the expressions above see Gagne and Wagner 1998 In Eqs 24 25 we have used the effective number of images N f instead of the actual number of images This is because for each imaging condition there are M files of 32 frames each used in the calculation but the 32 frames in any of the files are not necessary statistically independent because of lag The lag effect may depend on the spati
4. The average image for the signal the catheter located in the phantom b The average image for the background the catheter removed from the phantom c A one frame sample from the image series with the catheter in the phantom 65 kV 5 7 mA filtration 2 1 mm Al 0 25 mm Cu EAK 3 93 mGymint SNR 891 s SNR of single frames 10 7 rate 40 STUK A196 Table 1 shows the measured imaging conditions together with the measurement results of SNR x and dose rate Note that the data have not been acguired by using the automatic dose rate control and the dose rate does not therefore decrease with an increasing tube voltage This does not affect our analysis here becausethe system is well quantum noise limited this can be seen e g from the data at 2 1 mmAI filtration 65 kV 0 3 and 2 1 MA Therate of the effective dose was calculated with the PCXMC program Tapiovaara et al 1997 The effective dose data given in Table 1 correspond to the PA projection in a cardiological procedure of an obese patient 174 cm height and 105 kg mass Figures 7 and 8 show the doseto information conversion factors SNR x dose rate calculated from the data and corresponding to the air kerma rate and the rate of effective dose respectively The data have been plotted as a function of the x ray tube voltage for both filter choices Table I Measurement data for the optimisation example of detecting a catheter in the heart of a 25 cm thick patient
5. can be seen in figures 3 and 4 which represent the estimates of SNR in single frames and the SNR te of the image sequences respectively The data have been measured for 1 cm diameter PMMA disk details of three thicknesses 0 mm 0 9 mm and 3 mm thick the zero thickness corresponds to no actual signal in the image true SNR 0 The same phantom and geometry as described above was used n the measurements the anti scatter grid was used the x ray tube voltage was 72 kV and the tube current 0 7 mA 4 Analytical de biased 100 SNR 2 No signal N L Template method x uu SNR 2 No signal Analytical de biased SNR 2 0 9 mm signal 0 9 mm detail Template method SNR 2 0 9 mm signal Analytical de biased SNR 2 3 mm signal gt lt Template method SNR 2 3 mm signal SNR 2 true 0 9 mm signal SNR 2 true 3 mm signal SNR estimate 0 1 10 100 1000 Number of image files M Figure 3 The SNR of single image frames estimated by two different methods for three 1 cm diameter PMMA disks of various thicknesses 0 0 9 and 3 mm The dashed curves correspond to analytical de biased estimates Chapter 3 7 and the continuous curves to the template method Chapter 3 9 The horizontal continuous lines without data points show the expected unbiased SNR value for the 3 mm and 0 9 mm detail the latter calculated by scaling the SNR of the 3 mm
6. especially in the case of unknown anatomical background the variability does not necessarily conform with the underlying assumptions of NPS analysis noise stationarity and ergodicity 12 STUK A196 2 3 Summary measures of image quality SNR NEQ and DQE Presently image guality assessment in medical imagingis most often based on the statistical decision theory or signal detection theory SDT and uses quantities such as the ideal observer s signal to noise ratio SNR joa noise equivalent quanta NEQ and detective quantum efficiency DOE The applicability of this approach for several medical imaging modalities was summarised by Wagner and Brown 1985 and has been reviewed in ICRU 1996 In digital imaging systems images can be easily manipulated e g their brightness and contrast can be changed and images can be spatially filtered to suppress noise or improve sharpness Therefore the factors MTF NPS and contrast transfer are not of much use in digital imaging if one of them is used alone without reference to the others For example the MTF can be adjusted to almost any shape by filtering the image Such filtering affects also the NPS however and therefore a summary measure combining these factors properly such as SNR NEQ or DQE is required for describing the system performance The same applies to contrast which can be manipulated in electronic imaging systems to an arbitrary degree but affects both the signal and the nois
7. the FluoroOuality20 exe file is most conveniently installed in the folder X Koe In this case it is handy to delete the subfolder mageData then the data search will begin directly in the data folder used by these acquisition programs X Koe 52 APPENDIX A FLUoROOuaLITY v 2 0 User s GUIDE STUK A196 Using the program Buttons When started the program displays the Analysis form below The form has nine buttons and other functions TTO if mai igual zi ore ic Pee nni m Han Vumrmpaaiinp mun Ji i 1m ase MT Lin nr mi Bii k hormi ze rar ja 1 F ii pial sil 2 Skibi F ACT IL AE TL on Tirer ima are peered ip tishini har m iia GY am n harim ni ama iuris s Peepers yj enk rT e a Eii HPS in nm z pom pro o TEN mm mmm Lu m Fainal pas 11 A EX vd at it Wien ALOT Aaji ii 53 STUK A196 FLUOROQUALITY v 2 0 User s Guide APPENDIX A The Compute data button is used for analysing the image data computation of results When clicked an open dialog window is shown where the user may choose one or several image data series indicated by their parameter file name xyzxyzPM dat Several files can be chosen by using the Ctrl or Shift keys When clicked OK the program calculates the analysis data for all of the chosen image series If the Calculate spatiotemporal NPS checkbox is checked the program calculates these data as well Calculating spatio tem
8. 1 M Signal case data first 60 STUK A196 APPENDIx D THE FILES AND FILE FORMATS REQUIRED FROM THE ACOUISITION PROGRAM The acquisition program must produce a set of files whose contents are described below The measurement data consist of an acguisition parameter file M image files for the signal and M imagefiles for the background The data for each separate measurement is identified by an arbitrary six letter name which is being denoted below as xyzxyz The acquisition parameter file PM DAT The acquisition parameters are stored in file whose name is xyzxyzPM dat replace xyzxyz by the name you wish to use This file must contain seven data lines The first line must contain an integer number intended to show the opening mode of the frame grabber board if there are variable modes of operation this datumis not actually used in the program but is printed in the output data file xyzxyz txt only After the datum the row may contain also an explaining text that is not being used in the program In order to ensure compatibility with future versions of FluoroQuality we suggest that the number O is input here e Thesecond line must contain the actual duration of each image sequence in centi seconds For example the duration of the 32 frames in a 50 HzTV chain is 32 40 ms 128 cs After the datum the row may contain also an explaining text that is not being used in the program e Thethird line must contain the number of r
9. Bias issues of SNR measurements As already discussed in chapter 2 3 an uncorrected measurement of SNR from the average image and NPS data results to biased results mainly because the residual noise in the averaged data will be interpreted as being part of the signal On the other hand the residual noise in the average images that are used in the DCsHFs template method of SNR measurement Chapter 3 9 causes that the results from this method are ow biased in this case the noisy template does not accurately correspond to the actual signal and doesn t therefore perform as well as a noiseless template would do The SNR and SNR te estimates from three measurement methods i the analytical biased estimate ii the analytical de biased estimate and iii the estimate from the DCsH F s template method are shown in Figs 1 and 2 for 0 9 3 7 mm thick PMMA disk signals of 1 cm diameter obtained at different imaging conditions varying the dose rate the optical aperture between the image intensifier and the TV camera and the use of an antiscatter grid The background phantom used in the measurements consisted of a 2 mm copper plate at the x ray tube housing and a 4 7 cm PMMA block placed on the x ray image intensifier The x ray tube voltage was 72 kV the source to i mage distance 108 cm and the x ray beam size at the x ray image intensifier entrance plane 23 cm x 23 cm The number of image files M was 40 in these measurements i e 40 image seq
10. SNR is calculated as the average of two estimates Chapters 3 7 and 3 10 rate Total filtration kV mA Air kerma rate Effective SNR mGy min dose rate 1 s nSv min 2 1 mmAl 45 6 3 11 0 86 5 574 2 1 mmAl 50 2 5 5 75 56 2 417 2 1 mmAl 55 1 0 2 84 33 2 273 2 1 mmAl 60 0 5 1 70 23 2 165 2 1 mmAl 65 0 3 1 18 18 3 121 2 1 mmAl 65 2 1 8 55 133 792 2 1 mmAl 70 1 1 5 15 90 0 466 2 1 mmAl 75 0 7 3 68 71 8 331 2 1 mmAl 80 0 5 2 90 62 6 260 2 1 MMAI 85 0 5 3 22 76 1 249 2 1 mmAl 90 0 5 3 55 91 0 280 2 1 mmAl 95 0 5 3 90 107 290 2 1mmAl 0 25 mmCu 50 6 3 1 42 31 8 310 2 1MMA1 0 25 mmCu 55 6 3 2 22 58 8 531 2 1MmMA1 0 25 mmCu 60 6 3 3 20 97 3 747 2 1MMA 0 25 mmCu 65 5 7 3 93 134 891 2 1MMA 0 25 mmCu 70 3 7 3 36 126 693 2 1MMA1 0 25 mmCu 75 2 3 2 61 107 505 2 1mmAI 0 25 mmCu 80 15 2 09 92 4 350 2 1MmMAI 0 25 mmCu 85 1 1 1 84 86 5 305 2 1mmAl 0 25 mmCu 90 0 8 1 56 77 2 209 2 1mmAl 0 25 mmCu 95 0 6 1 39 71 8 176 41 STUK A196 20 15 pri SNR rate EAKrate 1 uGy 10 N ER Egg x 0 40 60 80 100 X ray tube voltage kV Figure 7 The dose to information conversion coefficient for detecting a 13 mm long piece of the USCI 5F catheter in a 25 cm thick patient In this figure the entrance air kerma rate free in air is used x total filtration 2 1 mm Al D total filtration 2 1 mm Al 0 25 mm Cu curves fitted by hand 800 600 400 X 200 SNR
11. frequency u 31 lt u lt 32 corresponds to the spatial frequency u X and the 2D NPS values can be obtained by multiplying the values calculated in FluoroQuality by XY Thetemporal length of the 32 frame image sequences is denoted as T 22 STUK A196 One should also note that all values in FluoroOuality are calculated in terms of the pixel values If the user so wishes these values can be later converted to correspond to some other guantities e g x ray fluence by applying the proper conversion factors Such a conversion makes a difference only in the numerical values of the average images and NPS data the signal to noise measures are not affected 3 2 Average images FluoroOuality calculates the average signal and background images recorded with and without the signal detail in the phantom respectively as 10 zi j syle i j k m s 40 1 m 1 k 1 The average images for the signal and background cases are shown on the FluoroQuality display form 3 3 The net signal and its frequency spectrum The net signal is obtained as 11 Agli j gii j oli j and the spatial frequency spectrum of the signal is defined as 12 DG e v Deli I The net signal image and the signal spectrum are shown on the F luoroOuality display form The relationship between this signal spectrum and the one used in eg 7 is similar tothe relationship between the noise guantities R and W in eg 14 23 STUK A19
12. in linear systems the gain K the modulation transfer function MTF and the noise power spectrum NPS symbol W These measurements are then combined to obtain the noise equivalent quanta NEO the detective quantum efficiency DQE or the ideal observer s signal to noise ratio SNR _ for a specified signal As ideal The NEQ of a linear imaging system is defined as K MTF f f Wats Spatial frequency f is expressed here by its horizontal and vertical components f and f because images are two dimensional objects NEO can be interpreted as the number of quanta actually photon fluence at the input of a perfect detector that would yield the same output noise as a function of spatial frequency as the real detection system under consideration In other words NEQ expresses the quality of the image data by the photon fluence that the image is worth at each spatial frequency By comparing the NEQ with the actual photon fluence Q used for forming the image one obtains the DQE 2 NEQU i 5 NEQ f gt f Q which can be interpreted as expressing the efficiency with which the imaging system has utilised the available photons for a perfect system DQE 1 for all Spatial frequencies DQE expresses therefore rather the quality of the equipment and the efficiency of radiation use than the quality of the image itself a low dose radiograph is bound to be noisy and therefore of not high quality although the DQE ma
13. lag are given in the output of FluoroQuality One is the lag measure related to the comparison of SNR in a single frame and SNR of the ideal observer rate SNR neat debiased SNR rate ideal debiased 30 Lag pwur The other one is calculated similarly to the DC HF suppressing observer and the third one is obtained from the DCsHF s observer by excluding the spatial frequency axes u O or v 0 from the summation S AG u v YR u v R u v u 0 v4 0 YAG v R yin g Us V FR gi UV u 0 v 0 31 Lag DCsHFs excl axes This last lag measure was developed in the past when we often experienced excessive noise on the spatial frequency axes cross like shapein the NPS We have not seen these artefacts with NPS calculations made with FluoroQuality however Anyway in our experience this last lag measure is the most stable of the measures presented above and it is used in calculating the SN R ste cours AS will be explained in 3 10 The error 1 STD in the lag measure is estimated from the expected precision of the average images and the NPS 3 9 The DCsHFs observer s SNR obtained using the template method In addition to the analytical calculations explained above FluoroQuality measures the DCsHF s observer s SNR in the single frames by the following method A template is calculated from the net signal Ag i j by first calculating the average pixel values of its even and odd pixel rows these correspo
14. land Dekh Pro pcie Alora yA pe SNR ee pe om gt yi jr TT lamin 1 58 He 234 He 313 He 3 51 He 5 47 Hz 7 03 Hz TI Hz 539 Hz 102Hz 103 Hz 11 7 Hz 125 Hz sae 0781 Hz 155 Hz Pami 313 Hz 17 Hz 163 Hz SLT Hz PISHI TA Ha a Sibir aH Ha 10h Wi ii 125 Hr The spatio temporal NPS values are in terms of pixel values and the normalisation does not indude the spatial and temporal extent of the images N ominally a 1 mm area and 1 s duration is assumed to normalize properly to units of mm s the data should be multiplied by the area of the images in mm and the seguence s temporal length in s Thetemporal freguency shown above the spectra is calculated as based on the image duration datum that is given in xyzxyzP M dat TheNPS of any image data can be displayed the data name is shown in the form caption Use the Open button The NPS form is closed and the basic form displayed by clicking the Close button 57 STUK A196 APPENDIX B INTERPRETATION OF THE DATA IN XYZXYZ TXT Thefilefirst contains identification information Data name and details of the acquisition of the image data The next section is a summary of the analytical calculations from the average image and NPS measurements in the basic form The biased SNR estimates are calculated for the DCSHFs observer and the PWMF ideal observer and the de biased estimate for the PWM F observer The table shows the results as a function of included frequency ch
15. many errors false positives and false negatives the observer makes The less detection errors the observer makes the better the image quality is This performance can be summarised by the observer s SNR at the decision stage it describes the accuracy of the the observer in classifying images with and without the signal to the correct signal and background classes It would not be of much interest to study how an unskilled or inefficient observer would succeed in detecting the signal the results would describe more the observer s in ability than the actual information in the images In order to get a unique performance figure which describes the actual quality of the image data in an absolute scale one uses the best possible observer the ideal observer for observing the images This observer uses all the information in the images and all available prior information in the optimal way to make its decision The ideal observer then achieves the lowest detection error rate that is possible by using the image data Therefore the performance of the ideal observer is a measure of the amount of information in the image which is relevant to the specified imaging task In an SKE BKE task any decision errors that the ideal observer makes result from the image quality not being perfect full prior data is given to the observer One must be cautious in interpreting the SKE BKE data however because sometimes the detection task may become too tightly speci
16. neli n kertym nopeus SNR te spatiotemporaalinen kohinan tehospektri NPS kuvan hitaus optimointi laadunvarmistus mittausmenetelm tietokoneohjelma Tiivistelm Raportti kuvailee STUKissa kehitetty FluoroQuality nimist tietokone ohjelmaa jota voidaan k ytt l ketieteellisten l pivalaisulaitteiden kuvan laadun mittaamiseen ja analysointiin Mittausmenetelm perustuu tilastolli seen p t ksentekoteoriaan ja keskeisin mittaustulos on signaali kohina suhteen neli n kertym nopeus SNR te Kuvadatasta analysoidaan my s muita suureita kuten yksitt isten videokuvien signaali kohinasuhde SNR spatiotemporaalinen kohinan tehospektri ja kuvan hitaus lag Mittaus menetelm voidaan k ytt esimerkiksi l pivalaisukuvan laadun ilmaisemi seen l pivalaisun kuvanlaadun ja potilaan annosnopeuden optimointiin sek l pivalaisulaitteiden aadunvarmistukseen Raportissa on katsaus mittausten taustalla olevaan teoriaan ja eri suureiden mittausmenetelm t on selostettu Esimerkkin n ytet n menetelm n k ytt er n l pivalaisututkimuksen optimoinnissa Ohjelman k ytt ohje on raportin liitteen Ohjelman k ytt lisenssin saa tekij lt ilmaiseksi tutkimustarkoituksiin ja mittausme netelm n k ytt kel poisuuden arviointiin STUK A196 Contents Abstract Tiivistelm 1 2 Introduction Background 2 1 General concepts of image guality 2 2 Basicfactors of image guality MTF contrast and NPS
17. the noiseis analysed from image samples obtained from the same location of the image receptor after subtracting the actual averaged image from the samples Background variability is then treated as being a deterministic known structure which does not impair detail detectability This may not always be realistic for a human observer who may in some cases suffer from background variability more than from actual stochastic noise Kotre 1998 Bochud et al 1999 Burgess et al 2001a and b Marshall et al 2001 but is certainly applicable to the ideal observer Human observers seem to operate somewhere between the two interpretations background variability appears to function as a mixture of noise and deterministic masking components For a more detailed discussion on this matter see e g Burgess et al 2001b and the references therein In FluoroQuality the measurement of signal transfer characteristics is not attempted and only the visibility of static details is considered Instead of using a model of the signal and its transfer the mean detail image is obtained directly as the difference of averaged almost noiseless images that are acguired with and without the signal detail in the phantom This difference image automatically contains all thefactors that affect either image sharpness or contrast As already stated above the spatio temporal NPS is measured however and is displayed as 2D cuts at different temporal freguencies N However
18. 2 3 Summary measures of image quality SNR NEQ and DQE 2 4 Visual assessment Theimage guality guantities measured with FluoroOuality 3 1 Notation and conventions 3 2 Average images 3 3 Thenet signal and its frequency spectrum 3 4 TheNPS of individual image frames and the spatio temporal NPS at zero temporal frequency 3 5 The full spatio temporal NPS 3 6 TheSNR measures obtained by integration analytical biased data 3 7 TheSNR measures obtained by integration analytical de biased data 3 8 Temporal lag 3 9 TheDCsHFs observer s SNR obtained using the template method 3 10 SNR te 3 11 Bias issues of SNR measurements SNR e and detail visibility An example of using FluoroQuality for imaging technigue optimisation Acknowledgements References and further reading APPENDIX A FluoroOuality v 2 0 User s Guide APPENDIX B Interpretation of the data in xyzxyz txt APPENDIX C Which files are necessary to keep and the contents of the datafiles APPENDIX D Thefiles and file formats reguired from the acquisition program oon sf WU 13 20 22 22 23 23 24 25 25 26 28 28 30 30 36 39 45 46 51 58 59 61 STUK A196 1 Introduction This report describes FluoroOuality a computer program that is intended for the measurement of physical image guality in medical fluoroscopic eguipment The measurement method of the program is based on statistical decision theory SDT and fluoroscopic image qual
19. 6 3 4 The NPS of individual image frames and the spatio temporal NPS at zero temporal freguency The variance at each spatial freguency R u v which is related to the two dimensional NPS see e g Tapiovaara and Wagner 1993 is measured as a F le i Jk al k m 3 ee i 13 R u v We j lk i j k m n The two dimensional NPS corresponding to individual frames is then calculated as 14 Wy u X v Y T u v The measurement is made separately for both the signal s 1 and background images s 0 When reporting the NPS FluoroOuality uses the value 1 for both factors X and Y If the user wishes to normalise his her data to the actual size of the image the calculated NPS values should be multiplied with the measured value of XY and the spatial frequencies are obtained by dividing the displayed integer values of u and v by X and Y respectively It is again emphasized that the normalisation in 14 differs from some other texts because of the DFT normalisation used the resulting NPS is the same however Similarly the variance of the summed 32 frame long sequences corresponding to the zero temporal frequency component of the 3D spatio temporal NPS is measured as N 2g l A g ijk my M m 6 UI H 1 E NN ol 15 R ms W v A F 6A g i j k m M 1 n and the corresponding zero frequency component of the 3D spatio temporal NPS of the sequences is calcula
20. NPS for the signal and background data sets are saved in files AnalysedData xyzxyzsg spt and AnalysedData xyzxyzbg spt The files contain the minimum and the maximum value of the data the temporal sequence length and the NPS values as eight byte real numbers The data for each temporal frequency are written line by line each temporal frequency data starting with the spatial 0 0 frequency and the data are written for each positive temporal frequency The file AnalysedData xyzxyzSN Rsp dat contains the spatial freguency components of the de biased SN R spectrum of single frames These data are shown in the SNR one frame image on the basic form display Thefile contain the values as eight byte real numbers line by line starting from the 0 0 frequency The file AnalysedData xyzxyzSN ROsp dat contains the spatial frequency components of the de biased SNR spectrum These data are shown in the SNR2 0 Hz image on the basic form display The file contains the values as eight byte real numbers line by line starting from the 0 0 freguency The text file AnalysedData xyzxyzSNR1280 dat contains the decision variable values of the direct one frame SNR measurement the values refer to imagefiles 1 M and are given for each frame Signal case data first The text file AnalysedData xyzxyzSNRSE 01280 dat contains the decision variable values of the direct averageframe SNR measurement method the values refer to image files
21. NR ve aS depicting the effective time that a human has integrated the information from a live image seguence This interpretation should not be taken literally however the detection times in the 16AFC tests above were very long from about 30 s to several minutes for each single observation Anyway the obtained result of t a1 3s may be a useful rule of thumb for interpreting the meaning of SNR x from a human observer standpoint This rule seems to apply even to the SNR threshold tests mentioned above the perceived SNR thresholds Via NE saa range from 4 4 to 6 9 and are in good agreement with values guoted for static images It is however noted that we have earlier Tapiovaara 1997 found a relatively high efficiency for human observers detecting static low contrast details in dynamic noise These earlier measurements were made using the 2AFC method and finitelength image sequences which were replayed in a continuous loop nterpreting the data from that study in a similar way as is done here would suggest an effective information integration time of at least 0 8 s given that the display contrast is optimal We do not presently know the reason for this difference between our present and earlier data but suspect that it could be due to the much easier decision task in the 2AFC experiment compared to the 16 AFC experiment used here having only two possible signal locations near each other does not place similar demands on memory as
22. O 5 TUK STUK A196 May 2003 OBJ ECTIVE MEASUREMENT OF IMAGE QUALITY IN FEUOROSCOPIC X RAY EQUIPMENT FLUoROQUALITY M Tapiovaara STUK SATEILYTURVAKESKUS Osoite Address Laippatie 4 00880 Helsinki STRALSAKERHETSCENTRALEN Postiosoite Postal address PL PO Box 14 FIN 00881 Helsinki FINLAND RADIATION AND NUCLEAR SAFETY AUTHORITY Puh Tel 358 9 759 881 Fax 358 9 759 88 500 gt www stuk fi The conclusions presented in the STUK report series are those of the authors and do not necessarily represent the official position of STUK ISBN 951 712 688 3 print ISBN 951 712 689 1 pdf ISSN 0781 1705 Dark Oy Vantaa 2002 Sold by STUK Radiation and Nuclear Safety Authority PO Box 14 FIN 00881 Helsinki Finland Phone 358 9 759 881 Fax 358 9 7598 8500 STUK A196 TAPIOVAARA Markku STUK A196 Objective Measurement of mage Quality in Fluoroscopic X ray E qui pment F l uoroOuality Helsinki 2003 50 pp apps 13 pp Keywords medical imaging x ray imaging fluoroscopy image quality statistical decision theory ideal observer quasi ideal observer detectability signal to noise ratio SNR accumulation rate of the signal to noise ratio squared SNR ja Wiener spectrum spatio temporal noise power spectrum NPS temporal lag optimisation quality control measurement method computer program Abstract The report describes FluoroQuality a computer program that is developed in STUK and used f
23. al frequency and therefore the effective number of images at each spatial frequency is estimated separately as being RON ER Ro sum f t Ri sum f In FluoroQuality N is subjected also to the condition M lt N lt 32 M The above expressions are almost similar for the calculation of the 26 Ny f M 32 SNR te 27 SN Ree ideal s wa SNR f seg debiased where AG u v 28 SNR ae 3 Ran E 4 M 17 4 SNR o debiaca SNR SNR Bs A 29 o f seq abies M f seq debiased UM ET and M denotes the number of recorded signal or background image seguences There is no need to estimate an effective number instead of M because the seguences can be assumed to be statistically independent Typically the noise at low spatial frequencies is dominated by quantum noise which is affected by the lag in the imaging system The noise at high spatial frequencies is often dominated by temporally white electronic noise 27 STUK A196 3 8 Temporal lag Temporal lag was evaluated in terms of the lag factor F in Tapiovaara 1993 This factor compares the information rate in an image seguence to the information in a single frame and shows the effective number of independent image frames per unit time In that paper F was defined in terms of the DC suppressing observer Here lagis reported in terms of guantities related to 1 F and the unit of the lag measures is s Three slightly different measures of
24. annels all frequencies the frequencies below 30 f in below 15 f_ and below 7 f_ included Here f_ isthe spatial frequency resolution in the measurement f an 1 X and X is the spatial extent width or length of the image area analyzed The lag factors have been calculated analytically from the average images and the NPS data Here they are reported in units of a second and are the reciprocals of the lag factors specified in Tapiovaara 1993 where the lag factor was defined as the number of statistically independent frames in a second The direct SNR measurement section contains the SNR data that are obtained by employing the DCsHFs observer The SNR x data have been obtained by two ways The first figure is the SNR of a single frame divided by the lag from the analytical calculation spatial frequency O axes excluded from the calculation The second is the result from direct measurement using the averaged sequence data The latter might be preferred by being more direct but may suffer from the small number of image data M V Therefore we prefer using the first SNR measure The user is warned if the variance of the DCsHFs observers decision variable differ by more than 20 in the signal included and signal absent cases In principle the variances should be equal The SNR inge frame Measurement has an extra check for the normality of the decision variable the chi test There is a warning if the data are not compatible with th
25. asks 36 STUK A196 detectability with the condition of 50 correct responses in an 18 AFC test which corresponds to d 1 78 Using the estimate of F 50 this would correspond to the detectability of signals with SNR gt 2 5 In fluoroscopy the threshold contrast i e the lowest contrast detail that the observer subjectively judges as perceivable is likely to depend on several factors in addition to the SNR e These factors may include e g the instructions given to the observers the design of the test object the displayed contrast the properties of image noise the allowed observation time and any background non uniformity One should also note that the inter and intraobserver variability in visibility threshold tests is large the threshold is difficult to define and keep and may therefore have a different meaning to different observers and to a given observer at various times In the human observer tests that we have made we have found an average SNR x for the observers declaring a detail in a noisy fluoroscopic image as just visible tobe around 60 s This threshold SNR x did not seem to be independent of the noise or dose level however but increased with the dose rate i e decreasing noise the average threshold SNR x was 87 s for a two fold dose rate and 140 s for a four fold dose rate compared to the lowest dose rate in the tests So the observers visibility threshold did not impro
26. average signal image minus average background image Thesguareof the FFT of the net signal e Thespectrum of theideal observer s prewhitening matched filter SNR for the one frame and temporal zero freguency cases The display is for the de biased estimate unless otherwise indicated by a blue text Biased SNR data under the images This happens if the files containing the SNR spectra have been deleted The text in the memo is identical to the file AnalysedData xyzxyz txt and displays summary information mainly on SNR measurements Show NPS button The Show NPS button is used for displaying the spatio temporal NPS spectra of the signal and background image set The form used for NPS display is shown below The minimum and maximum value of the data set is shown in the upper left corner The brightness and contrast of the NPS images can be set by modifying the Min and M ax edit boxes these values set the levels of black and 56 APPENDIX A FLUoROOuaLITY v 2 0 User s GUIDE STUK A196 white intermediate grey levels are used for the values between Min and Max Checking the L og box results in displaying the logarithm of the NPS values this helps in providing a wider dynamic range for the display In the NPS form modifying the values in the M in or M ax editboxes or checking thelog box does not result in immediate display of the NPS in order to display the data the user must click the Draw button Lo Png pn Edi oe
27. cknesses in various imaging conditions Number of signal and background image files M 40 The data are plotted against the true SNR of the details rate In the figures it is seen that the uncorrected biased SNR estimates suffer increasingly from the bias when the SNR gets smaller therefore these results are not of much use The template method esti mates are somewhat low biased at very low SNR Ss as expected and the de biased SNR estimates seem to get slightly high biased at low SNR s in spite of the bias correction The de biased analytical SNR data include also a few outliers whose valueis notably higher than expected These outliers are characterised by their signal spectra including more power in their high spatial frequency region than would be expected from the actual low frequency signal used it seems that in some cases the averaging process of the signal and background images has not cleaned image noise in the average images as well as would be expected by the number of images These outliers stood out also in the analytical SNR calculation results by the strong dependence of their SNR values on the maximum included frequency fi A possible method of reducing the bias of truly low frequency signal details is then to use the SNR estimates based on 32 STUK A196 a sufficiently low value of f in the analytical calculation Gagne and Wagner 1998 The dependence of the bias on the number of analysed image files M
28. datum 33 STUK A196 1000 3 mm detail y 100 0 9 mm detail 10 SNRrate estimate 1 5 10 100 1000 Number of image files M Figure 4 The SNR e estimated by two different methods for three 1 cm diameter PMMA disks of various thicknesses 0 0 9 and 3 mm The dashed curves correspond to analytical de biased estimates Chapter 3 7 and the continuous curves to the template method Chapter 3 10 The horizontal continuous lines without data points show the expected unbiased SNR for the 3 mm and 0 9 mm details the latter calculated by scaling the SNR ofthe 3 mm datum rate For darity of illustration the analytical biased SNR estimates have not been plotted in these figures they were again positively biased to such a degree that they were almost useless The bias of the analytical de biased estimate increases slightly with a decreasing number of image files the positive bias of this SNR estimate seems to be independent of the signal strength and is of the order of 3 4 when M 10 The bias of the template based SNR2 estimate is negative as expected and increases with a decreasing M This bias is smaller for the zero and 0 9 mm details than the bias of the de biased estimate but larger although small for the 3 mm detail The disagreement between the de biased SNR estimate and the template method SNR estimate of the 0 9 mm detail is in accordance with Figure 1 and suggests that even M 160 ma
29. ded task well Image quality then becomes a task dependent quantity images ranked by one imaging task will not necessarily rank similarly in another task For example if the visibility of small sized details is important imaging system performance at high spatial frequencies may be amore important factor than imaging system performance at low spatial frequencies and vice versa if the visibility of large low contrast objects is required ICRU 1996 It could be thought that such a definition of image quality would obscure matters image quality is then not solely dependent on image properties but also on the detection task the observer s a priori information on the task and the observer s ability to use both the prior information and the image STUK A196 information for his decisions This apparent difficulty cannot be avoided but can be dealt with by specifying thetask and the observer in detail Theproblem of prior information is commonly treated by considering the case of full a priori information the observer is given all information on the expected signal and background the signal transfer properties of the imaging system and the properties of the image noise The only thing that the observer does not know a priori is whether the signal is in the image or not the observer s task is to make a decision on the signal s presence The detection experiment is repeated many times and image quality is measured statistically by observing how
30. e In the SDT framework image quality assessment applies to the image data stage and describes the performance of a specified mathematical observer when it analyses images The observer calculates a decision variable which describes the observer s confidence for the presence or absence of the specified detail in the image an image will be denoted by the symbol g Detection performance is measured statistically on an ensemble of images and is described by the separability of the conditional distributions of the decision variable D g s for the signal and background cases The overlapping of these distributions specifies the probability of both types of detection errors false alarms and misses and can be presented by the observer s receiver operating characteristic ROC curve Frequently this separability of the distributions is reported in terms of the separation of their means divided by their standard deviation this is the observer s signal to noise ratio D g signal D g background 1 SNR Op The SNR description is sufficient for specifying the observer s performance when the conditional distributions of the decision variable are normally distributed and have equal variance for both the signal and background cases 13 STUK A196 For the evaluation of the physical or technical guality of images or imaging systems the ICRU report 1996 suggests the measurement of the large scale system transfer function
31. e image are written in M image data files whose names must be xyzxyz01 ims xyzxyz02 ims xyzxyz40 ims In this example M is 40 The image data corresponding to the signal absent case are written in files xyzxyzOl imt xyzxyz02 imt xyzxyz40 imt In this example M is again 40 The pixel values are written line by line as bytes beginning from the upper left corner of thefirst imagein theseguenceuntil all lines and all frames have been written In Fortran the code for writing the pixel values g i j K of a sequence in the file xyzxyz01 ims is INTEGER 1 g 64 64 32 NAME xyzxyz01 ims OPEN 1 FILE NAME STATUS NEW RECL 64 FORM UNFORMATTED DO 2 K 1 32 DO 2 J 1 64 2 WRITE 1 g 1 J K 1I 1 64 CLOSE 1 62 APPENDIX D THE FILES AND FILE FORMATS REQUIRED STUK A196 Here refers the horizontal location of the pixel J the vertical location and K the frame Note that in Fortran each file is started with a value 75 denoting the beginning of the file and ended with the value 130 denoting the end of the file Also the length of the physical record is written at the beginning and end of each record therefore the actual record length here is 64 2 66 bytes and the total size of each of the image sequence files is then 32 64 64 2 2 135 170 bytes In FluoroQuality written in Delphi Object Pascal the image files e g xyzxyz01 ims are read by NAME xyzxyz0l ims Assi
32. e intensifier and an XTV 11 video chain with a low lag Vidicon Newvicon pick up tube operated in the 625 lines frame interlaced scan mode A stationary antiscatter grid strip density 44 line pairs cm grid ratio 10 1 carbon fibre cover plates focusing distance 1 m was located in front of the image intensifier entrance plane The phantom was imaged in the fluoroscopic mode using various imaging conditions with and without an added 0 25 mm Cu filter several x ray tube voltages The x ray tube voltage and current were chosen manually but the gain in the video chain was automatically controlled The fluoroscopic sequences were recorded both with and without the signal detail 13 mm piece of the USCI 5F catheter in the analysed image area by using a Matrox Genesis LC frame grabber board and the NoiseAcquisition program see Appendix A The image sequences were analysed with the FluoroQuality program The N The optimum imaging conditions are defined as those that result to the lowest dose in the patient for a specified image quality 39 STUK A196 entrance air kerma rates EAK free in air were measured using a Radcal 9015 dosemeter equipped with a 10x5 6 ionisation chamber The x ray tube voltage was monitored using a Machlett Dynalyzer II voltage divider chain Figure 6 shows the average images and a sample image frame in one of the imaging conditions a b c Figure 6 An example of the images in the detection task a
33. e underlying assumption of the decision variable values being normally distributed If this warning is seen and the value is suspiciously high thereis probably a problem in the acquired data or other than purenoise processes in thefluoroscopy system These first two SNR estimates are usually high biased and should not be used as such for evaluating detail detectability 58 STUK A196 APPENDIX C WHICH FILES ARE NECESSARY TO KEEP AND THE CONTENTS OF THE DATAFILES If disk space need to be released and the spectral data need not be stored all other datafiles than xyzxyz txt where xyzxyz denotes the six character name of the data can be deleted This file contains the summary of the SNR data However other information is lost and the data must then be viewed using a text editor If the user wishes to keep the data needed for viewing them in the basic FluoroQuality form the files xyzxyz txt xyzxyzsg ave xyzxyzsg nps xyzxyzsg npl xyzxyzbg ave xyzxyzbg nps xyzxyzbg np1 and xyzxyzSN Rdataonform dat must be kept If the files xyzxyzSNRsp dat and xyzxyzSN ROsp dat are deleted the SNR spectra shown on the form will be the biased spectra this will be noted to the user by a text appearing below the spectrum e If the user wishes to be able to view the spatio temporal NPS the files xyzxyzsg spt and xyzxyzbg spt must be kept e If the user wishes to be able to view the original images and or to be able to recalculate the resul
34. ecorded seguences denoted as M in chapter 3 After the datum the row may contain also an explaining text that is not being used in the program The fourth line must contain two integer numbers The numbers represent the horizontal and vertical position of the analysed sub image The data are not actually being used in the program but are shown in the output data file xyzxyz txt only After the datum the row may contain also an explaining text that is not being used in the program The fifth line must contain two integer numbers The numbers represent the offset and white reference settings of the frame grabber board The data are not actually being used in the program but are shown in the output data file xyzxyz txt only After the datum the row may contain also an explaining text that is not being used in the program 61 STUK A196 THE FILES AND FILE FORMATS REOUIRED APPENDIX D Thesixth line contains a comment text that the user wishes to be shown in the output data file xyzxyz txt The seventh line contains the name of the acquisition program An example of the file xyzxyzP M DAT is shown below 0 7 opening mode dummy 128 7 cs collection time 40 js number of series 339 251 subimage location dummy 60 240 offset reference dummy This is an example of a comment line NoiseAcquisition The image files yy ims and yy imt The actual image data corresponding to the situation of the signal being in th
35. ency conditions would suggest 43 STUK A196 4000 3500 3000 2500 2000 1500 1000 500 0 T T 40 60 80 100 X ray tube voltage kV Maximum SNRrate at 6 3 mA 1 5 Figure 9 The SNR that could be expected at a tube current of 6 3 mA the maximum for the x ray equipment used in the measurement as based on the data in Table and the quantum noise limited behaviour of the imaging system x total filtration 2 1 MMAI p total filtration 2 1 mm AI 0 25 mm Cu curves fitted by hand Figure 9 shows an estimate of the maximum SNR te IN our x ray system at each tube voltage and filtration the maximum tube current in our x ray system is 6 3 mA The estimate is based on the data in Table and assumes that our imaging chain is quantum noise limited Now if the required image quality would correspond to an SNR te of 1500 s for example it would be better to image the patient without the copper filter because the copper filter would require a tube voltage of at least 80 kV in order to achieve the required image quality The efficiency is low at such high voltages and the aluminium filtered system operating at 55 kV would in this case allow an about 50 lower effective dose rate see fig 8 The actual optimum condition of this example would require a filtration between the two examples here the optimum filter would be the thickest filter that allows the use of an x ray tube voltage near 50 55 kV If we could use a higher pow
36. er practical implementation SNR estimation and bias reduction Proc SPIE 3336 231 242 1998 47 STUK A196 GagneRM MyersKJ and Quinn PW Effect of shift invariance and stationarity assumptions on simple detection tasks spatial and spatial frequency domains Proc SPIE 4320 373 380 2001a Gagne RM Boswell J S Myers KJ and Peter G Lesion detectability in digital radiography Proc SPIE 4320 316 325 2001b Goldman LW F luoroscopic performance tests using a portable computer frame grabber Wiener spectra measurements Med Phys 20 117 127 1992 Green DM and Swets JA Signal Detection Theory and Psychophysics J ohn Wiley and Sons New York 1966 Hanson KM Variations in task and the ideal observer Proc SPIE 419 60 67 1983 ICRU Report 41 Modulation Transfer Function of Screen Film Systems International Commission on Radiation Units and Measurements 1986 ICRU Report 54 Medical imaging the assessment of image quality International Commission on Radiation Units and Measurements 1996 Kotre CJ The effect of background structure on the detection of low contrast objects in mammography Br J Radiol 71 1162 1167 1998 Loo L ND Doi K and Metz CE A comparison of physical image quality indices and observer performance in the radiographic detection of nylon beads Phys Med Biol 29 837 856 1984 Marshall NW The practical application of signal detection theory to image quality assessment in x ray ima
37. er x ray system the optimum filter thickness would be larger This would bea high quality fluoroscopic image roughly discarding effects of resolution even the SNR e for a 1 mm piece of the catheter would be equal to about 1500 13 115 which suggests easy detectability of catheter details of such size A The lower filtration in this example is below the minimum allowed filtration of medical x ray equipment The use of such filters is not suggested a system working between the two filter examples considered would be the optimal choice here 44 STUK A196 6 Acknowledgements This work was partly funded by the European Commission s 5th Framework Programme 1998 2002 Nuclear fission and Radiation Protection Contract DIMOND II FIGM CT 2000 00061 This report is the sole responsibility of the author and does not reflect the opinion of the European Commission or STUK The European Commission or STUK are not responsible for any use that might be made of data appearing in this report 45 STUK A196 7 References and further reading Barrett HH and Swindell W Radiological Imaging Volumes and II Academic Press N ew York 1981 Barrett HH and Myers K Foundations of Image Science Mathematical amp Statistical Foundations J ohn Wiley and Sons to be published 2003 Beutel J Kundel H Van Metter R eds Handbook of Medical Imaging Vol 1 Medical Physics and Psychophysics SPIE Bellingham 2000 Bochud FO Valle
38. ersome however and the results depend not only on the image quality but also on the skills of the diagnosticians interpreting the images and the patient material Therefore the calibration of pati ent image based quality assessments is unclear and the results can hardly be accurately reproduced by others Simpler imaging tasks are thus required for the measurement and reporting of image quality in radiology One possibility is to use patient simulating phantoms and base the measurement on the detectability of phantom details that resemble important disease related structures in actual patients If the phantom is designed carefully it should be credible that the detectability of these details in phantom images is related to the detectability of important features in actual patient images and thus to the achievable accuracy in diagnostics 2 2 Basic factors of image quality MTF contrast and NPS Physical image quality depends on several factors The most important of these are image sharpness contrast and noise Other factors such as image distortions homogeneity and artefacts may be important too but are not treated here They are usually of less importancein conventional x ray imaging than the former and can be often corrected in the final image at least in principle In a sense image noise is the most important quality limiting factor in radiological imaging because it sets limits to the detectability of details and also restricts poss
39. es in the phantom may violate the assumption of noise stationarity and therefore make the NPS measurements uncertain Strong phantom structures may also make the detectability of the signal detail highly dependent on the actual location of the detail in the phantom and therefore achieving a good repeatability of SNR results may be difficult Therefore we recommend using a homogeneous phantom for most measurement applications For typical applications the signal detail should be strong enough for obtaining accurate results see Ch 3 11 but its contrast should be not too high in order to make the noise too much signal dependent or the SNR results dependent on the linearity of theimaging system and frame grabber The present version FluoroOuality 2 0 is compatible with the image data acquired using the program Filetall developed in STUK for the Data Translation DT 3852 frame grabber card or NoiseAcquisition developed by Olander and Sandborg in Link ping University Hospital for the Matrox Genesis LC frame grabber board These programs require the user to define a six character name for the data In this manual this name is denoted as Xyzxyz rate Reguirements for the data are given in Appendix D 51 STUK A196 FLUOROQUALITY v 2 0 User s Guide APPENDIX A Hardware and software requirements FluoroQuality runs under 32 bit Windows operating systems Windows 95 98 NT 2000 XP The memory requirement depends o
40. fied and not anymore correspond tothe actual detection task of interest An example of such a case is given by Myers et al 1990 The ideal observer is well known in the SKE BKE task when the image noise is signal independent and normally distributed and can be realised as a prewhitening matched filter Wagner and Brown 1985 In practical Y This task is often also referred to as the SKE BKE signal known exactly background known exactly task When there is less prior information on the detection task the ideal observer becomes mathematically more complicated An example of this is given in Brown et al 1995 where the case of unknown signal position was considered STUK A196 measurements it is not always necessary tousethis strictly ideal observer but one can be content of using a close approximation of it a quasi ideal observer which may be easier to implement in practice I n FluoroQuality image quality is assessed by estimating the SNR of both the ideal observer and a quasi ideal observer Tapiovaara and Wagner 1993 The above discussion is related to the image quality in general detection tasks In medical imaging the image is used as a means to get information of the health status of the patient and ultimately clinical image quality should be evaluated by the impact of the image to a correct diagnosis or tothe outcome of the treatment of the patient ICRU 1996 The evaluation of clinical performance is extremely cumb
41. ge intensifier TV fluoroscopy Phys Med Biol 46 1631 1649 2001 Marshall NW K otre CJ Robson KJ and Lecomber AR Receptor dose in digital fluorography a comparison between theory and practice Phys Med Biol 46 1283 1296 2001 Martin CJ Sharp PF and Sutton DG Measurement of image quality in diagnostic radiology Applied Radiation and Isotopes 50 21 38 1999 48 STUK A196 Metz CE Wagner RF Doi K Brown DG Nishikawa RM Myers KJ Toward consensus on quantitative assessment of medical imaging systems M ed Phys 22 1057 1061 1995 M oy P Signal to noise ratio and spatial resolution in x ray electronic imagers isthe MTF a relevant parameter Med Phys 27 86 93 2000 Myers KJ and Barrett HH Addition of a channel mechanism to the ideal observer model J Opt Soc Am A 4 2447 2457 1987 Myers KJ Rolland J P Barrett HH and Wagner RF Aperture optimization for emission imaging effect of a spatially varying background J Opt Soc Am A 7 1279 1293 1990 Pineda AR and Barrett HH What does DQE say about lesion detectability in digital radiography Proc SPIE 4320 561 569 2001 Siewerdsen JH Antonuk LE EI Mohri Y Huang W and Cunningham IA Signal noise power spectrum and detective guantum efficiency of indirect detection flat panel imagers for diagnostic radiology Med Phys 25 614 628 1998 Siewerdsen J H Cunningham A and affray DA A framework for noise power spectrum analysis of multidimensiona
42. ge sequences but the difficulty then is in defining the imaging time temporal lag spreads information to nearby image frames and the imaging time is not equal to the number of image frames multiplied by the nominal frame duration 18 STUK A196 SNR te PY Multiplying SNR ngierrame Py the noise lag factor F Tapiovaara 1993 see also Cunningham et al 2001 This factor is calculated from the spatio temporal NPS of the image seguence and expresses the effective number of independent image frames per unit time Because of lag this number is usually smaller than the framerate in the fluoroscopic system SNR te is calculated by both these methods in F luoroQuality The first method is more straightforward but may suffer from the imprecision caused by the small number of image sequences analysed In the image data system of FluoroQuality the number of analysed image frames is 32 times higher than the number of image sequences and we expect that the precision obtained with the latter method is better In addition to the aliasing problems discussed earlier these direct SNR measurement methods provide also a solution to a further problem in NEQ and DQE like quantities these latter quantities inherently apply only to imaging where a detail object affects only the intensity of the radiation and leaves the x ray spectrum behind the detail unchanged Tapiovaara and Wagner 1985 and 1993 Cahn et al 1999 In x ray imaging the detail of interest
43. gnfile D NAME FileMode 0 Reset D 66 for K 1 TO 32 do begin for J 1 TO 64 do begin BlockREAD D line 1 for I 1 to 64 do begin g 1 J K line 1I 2 end end end CLOSEfile D Here D is defined to be an untyped file g is an Array 1 64 1 64 1 32 of byte and lineis an Array 1 66 of byte 63
44. ibilities to get the details visible by image enhancement e g image sharpening and contrast increase mage noise is unavoidable in medical x ray imaging if the dose to the patient is to be kept low The sharpness of images is often evaluated visually by the resolution seen in line pair test object images The sharpness of linear shift invariant imaging systems can be better described by measuring the modulation transfer function MTF see e g ICRU 1986 The measurement may sometimes be 10 STUK A196 straightforward at least in principle but is usually complicated by problems caused by noise thelow intensity of theimage signal from thethin slit or small aperture used for the measurement and the wide dynamic range needed in measuring the line spread function or point spread function A further difficulty especially in electronic imaging is that the isotropy of the imaging system is not granted and the determination of the full two dimensional MTF may be necessary In fluoroscopy the situation is even more complex because time or temporal frequency constitutes a third dimension to the measurement and relates to the temporal blurring of the signal i e lag So far no practical methods for measuring the spatio temporal MTF of dynamic imaging systems have been presented in the literature The measurement of the MTF in digital imaging systems is further complicated by the fact that these systems are not necessarily shift invarian
45. image data The edit boxes show the displayed level of black and white with other grey values in between The brightness and contrast of the images can be adjusted by changing the values in the edit boxes The displayed image is updated immediately when a change in the edit box is made Checking the Log boxes result in displaying the image data as the logarithm of the actual value this may help in studying images with a large data range The black level is set at the logarithm of the value in the Min edit box and the white level at the logarithm of the Max edit box The displayed image is updated immediately when a change in the L og box is made When the cursor is moved on an image the position of the cursor and the actual image value not the displayed grey level is shown at the bottom of the form In addition to this the average of data in a 5x5 pixel ROI region of interest centered on the chosen pixel is displayed This average is denoted by Ave 25 Note that for the average images and the net image the position data show the coordinates of the pixel in the image in the Fourier transform images NPS FT of the net signal SNR spectra the position data relates to the spatial frequency component and the f 0 f 0 frequency is located at the centre of the image Spatial frequencies are expressed as integers The values of horizontal and vertical frequencies in units of mm can be obtained by dividing the displayed frequenc
46. in recognising jitter or background structure problems in the image data Notethat any image series can be chosen for viewing However the subtracted average is always from the image set identified in the caption of the form The Show NPS button is used for displaying the spatio temporal NPS data It is explained in more detail in the next chapter The Delete images button is an easy way to delete original raw i mage data when they are not needed anymore if the data have been analysed and there is no reason to directly view the images When this button is clicked an open dialog window is shown where the user can choose one or several data sets that are wished to be deleted files indicated by their parameter file name 54 APPENDIX A FLUoROOuaALITY v 2 0 User s GUIDE STUK A196 xyzxyzpm dat The file xyzxyzPM dat and the image files xyzxyzY Y ims and xyzxyzY Y imt will be deleted YY ranges from 01 tothe actual number of image files TheRedraw button redraws theimages on theAnalysis form A part of the images may occasionally become white because other windows eg OpenDialog may destoy their contents The images are not automatically updated but clicking this button restores them e The About button shows information of the program The Exit button is used for closing the program Other functional features of the analysis form On the left of each image on the form are shown the minimum and maximum values of the
47. ity is described primarily by the accumulating rate of the square of the signal to noise ratio SNR of the ideal and a quasi ideal observer s decision variable These SNR es relate to the detectability of a specified static detail in the image sequence by the ideal or quasi ideal observer The measurement is made by adding and removing the detail of interest in or from the phantom being imaged and analysing these recorded image data By choosing the phantom so as to sufficiently mimic the scattering and attenuation of radiation in a patient and choosing a detail that mimics a diagnostically important detail in the x ray examination the measurement should be closely related to the clinical quality of actual patient imaging By varying the phantom and the detail to be detected various detection tasks see also Hanson 1983 can be considered according tothe x ray examinations of interest In addition to SNR e the program produces also other quality related data which should be useful in evaluating the performance of the imaging system and for constancy testing The phantom need not necessarily be homogeneous in principle alsoan anatomic phantom may be used However high contrast phantom structures may violate the assumption of noise stationarity and therefore make the noise power spectrum measurements uncertain Such phantom structures may also make the detectability of the detail dependent on its specific location in the phantom and therefo
48. l images Med Phys 29 2655 2671 2002 Smith WE and Barrett HH Hotelling trace criterion as a figure of merit for the optimization of imaging systems J Opt Soc Am A 3 717 725 1986 Tapiovaara MJ SNR and noise measurements for medical imaging Il Application to fluoroscopic x ray equipment Phys Med Biol 38 1761 1788 1993 Tapiovaara MJ Efficiency of low contrast detail detectability in fluoroscopic imaging Med Phys 24 655 664 1997 Tapiovaara MJ and Wagner RF SNR and DQE analysis of broad spectrum x ray imaging Phys Med Biol 30 519 529 1985 Corrigendum Phys M ed Biol 31 195 49 STUK A196 Tapiovaara MJ and Wagner RF SNR and noise measurements for medical imaging A practical approach based on statistical decision theory Phys Med Biol 38 71 92 1993 Tapiovaara MJ Lakkisto M and Servomaa A PCXMC A PC based Monte Carlo program for calculating patient doses in medical x ray examinations Report STUK A139 1997 Tapiovaara MJ Sandborg M and Dance DR A search for improved technigue factors in paediatric fluoroscopy Phys Med Biol 44 537 559 1999 Tapiovaara MJ Servomaa A Sandborg M and Dance DR Optimising the imaging conditions in paediatric fluoroscopy Rad Prot Dosim 90 211 216 2000 van Trees HL Detection Estimation and Modulation Theory J ohn Wiley and Sons New York 1968 Wagner RF Low contrast sensitivity of radiologic CT nuclear medicine and ultrasound medical i
49. maging systems IEEE Transactions on Medical Imaging MI 2 105 121 1983 Wagner RF Characteristic images emerging from recent SPIE medical image symposia Proc SPIE 767 138 141 1987 Wagner RF and Brown DG Unified SNR analysis of medical imaging systems Phys Med Biol 30 489 518 1985 Wagner RF Beiden SV and Campbell G Multiplereader studies digital mammography computer aided diagnosis and the Holy Grail of imaging physics 1 Proc SPIE 4320 611 618 2001 Whalen AD Detection of Signals in Noise Academic Press New York 1971 50 STUK A196 APPENDIX A FLuoroOuaLITY v 2 0 User s GUIDE General FluoroOuality can be used for the measurement of the guality of fluoroscopic image seguences These seguences need first be recorded digitally by a PC frame grabber system using suitable software or reformatted from the digitally stored image seguence data of the x ray imaging system The FluoroQuality program then calculates and displays the SNR of single image frames the SNR te of the acquired image sequences the lag and the spatio temporal NPS for the acquired image data The program also displays the average image for the signal and background situations the net signal the SNR2 and SNR spectra and can be used for visualising the acquired image data The signal detail need not be located in a homogeneous phantom in principle alsoan anatomic phantom may be used H owever strong background structur
50. modifies also the x ray spectrum shape Therefore when optimising the x ray imaging conditions for example the x ray spectrum it is not sufficient to consider only NEQ or DQE but one must consider the spectral dependence of radiation contrast as well and include it in the factor AS f f above Spectral dependence is properly and automatically taken care of by the direct SNR measurement methods We make here one last note considering DQE The main application of this quantity is to describe the efficiency of theimagereceptor Therefore in the calculation of DQE the number of noise equivalent quanta is compared to the actual number of quanta impinging on the image receptor This is not directly the optimisation problem that is of interest in x ray imaging The efficiency in x ray imaging is better described by comparing the achieved image quality as related to the chosen task to the dose in the patient Therefore in many papers discussing optimal imaging conditions the optimisation process is based on maximising the efficiency of radiation usein terms of the dose to information conversion factor SN R dose or SNR J dose rate for example Tapiovaara et al 1999 Chakraborty 1999b This quantity helps in finding the most efficient conditions of imaging but even this is not sufficient by itself one also needs to decide on the image quality i e actual details and their detectability that is needed and to work at the lowest dose level at
51. n the operating system but is 64 MB RAM for a PC operated under Windows NT4 we suggest however a minimum of 128 MB The PC also needs sufficient hard disk space to store the image data about 10 3 MB for each measurement comprising of 40 background and 40 signal datafiles and the analysed data 1 37 MB for each measurement that the user wishes to keep In order to keep the computation times reasonably short one should use a fast PC 233 MHz is a suggested minimum clock frequency Installation of the program The program is installed by running the setup exe program of the FluoroQuality Installation Diskette The program can be installed in any folder The setup program creates also a subfolder AnalysedData which is used for saving the data that the program calculates Also a subfolder ImageData is created it is intended to be the folder where the image data are saved by the acquisition program When the analysing calculation is started the program first tries to look for image data in the folder ImageData if such a folder exists If not the file open dialog window is set to the program containing folder and the user locates the data him herself The setup program will add a shortcut to the FluoroQuality in the Programs menu For an even more easy access to the program you may create a shortcut to the FluoroQuality20 exe file on your desktop N If used with the above mentioned image data acguisition programs
52. nd to the even and odd video fields of one video frame and subtracting 28 STUK A196 these averages from the pixel values of the respective pixel rows of the net signal This results to a template whose DFT would have the value zero at frequencies 0 0 and 0 V a This template is calculated separately for each image seguencefile m by leaving this seguence out in calculating the average image and is denoted here as Ag por m i j This template is then cross correlated with each image frame in sequence m to get theDCsH F s observer s conditional decision variables 32 D ycsurs 8 km Yamnan i j 8 i j k m f ij The SNR is then calculated from these values as shown in Eq 1 There are also other possibilities to use the image data for the measurement For example one could divide the image data in two separate groups and use one group for estimating the averages and the other for testing the performance However this results to inefficient use of the data and unnecessary imprecision and bias in the results Another alternative would be not to leave out the image data being tested from estimating the average image this would lead to biased results Theerror 1 STD in the obtained SNR is estimated as bonr NO D DI 33 An J wa SNR pcs 0 Wte PI Ny 2 08 1 0 D10 0 DID where the effective number of images is estimated from T N M 34 LA pesnrs excl axes TheSNR measurement i
53. o frequency and at the Nyquist frequency the factor in the parentheses is N 2 N 1 for these frequencies FluoroQuality displays the de biased SNR2 spectra related to both the individual image frames and to the temporal zero frequency corresponding to SN R e and the detectability of a static detail in the fluoroscopic sequence to be explained later in the text The program 16 STUK A196 also reports the de biased SNR and SNR e calculated from Eq 8 the data are calculated as the sum over various freguency ranges Another possibility of measuring the SNR is not tousethe mathematical relationships between the SNR and its constituents Eqs 4 or 7 but to actually construct the mathematical observer and let it make decisions on the detail presence in images with and without the signal detail in the phantom This approach will be considered in more detail later If the measurement is done in either of these ways the result may depend on the exact position of the detail with respect to the pixel array if aliasing phenomena are present and thus vary somewhat from one measurement to another If necessary a solution to this is to make several SNR measurements with small shifts in the detail position and report the mean detectability Pineda and Barrett 2001 also discuss this solution In their simulations they found that a direct SNR measurement from the digital data is necessary when the signal size is of the order of or smalle
54. on of temporal frequency Goldman 1992 Tapiovaara and Wagner 1993 Tapiovaara 1993 Cunningham et al 2001 Siewerdsen et al 2002 It can be noted that the 2D spatial NPS of the individual images in the image sequence can be obtained by integrating the 3D spatio temporal NPS of the sequence over the temporal frequency In FluoroQuality both the 3D spatio temporal NPS and the 2D spatial NPS of single image frames are measured The 11 STUK A196 eguations for NPS measurement as used in FluoroOuality are given in Chapters 3 4 and 3 5 Thereareambiguitiesin noise measurements too For example any non homogeneity in the image background originating from the non homogeneity in the phantom or the image receptor is often considered as being noise This is reasonable if the observer does not know these structures and the structures actually vary from one image to another If the spatial variability stays constant in all images one cannot treat it as being random although the detailed structure of the non homogeneity would be unknown to the observer In practice the noise analysis is often made by subtracting a constant brightness value or a slowly varying fit from the image data before calculating the NPS This corresponds to considering other brightness variability as being noise and may result to false anomalous NPS values if the background structure does not change between analysed image samples In FluoroQuality another alternative is used
55. or measuring the image quality in medical fluoroscopic equipment The method is based on the statistical decision theory SDT and the main measurement result is given in terms of the accumulation rate of the signal to noise ratio squared SNR 4e In addition to this quantity several other quantities are measured These quantities include the SNR of single image frames the spatio temporal noise power spectrum and the temporal lag The measurement method can be used for example for specifying the image quality in fluoroscopic images for optimising the image quality and dose rate in fluoroscopy and for quality control of fluoroscopic equipment The theory behind the measurement method is reviewed and the measurement of the various quantities is explained An example of using the method for optimising a specified fluoroscopic procedure is given The User s Manual of the program is included as an appendix The program is available free of charge for research work and program evaluation purposes by contacting the author STUK A196 TAPIOVAARA Markku STUK A196 Objective Measurement of mage Quality in Fluoroscopic X ray E guipment FluoroQuality Hesinki 2003 50 s liitteet 13 s Englanninkidinen Avainsanat l ketieteellinen kuvantaminen r ntgenkuvantaminen l pivalaisu kuvanlaatu tilastollinen p t ksentekoteoria ideaalinen havaitsija kvasi ideaalinen havaitsija havaittavuus signaali kohinasuhde SNR signaali kohinasuhteen
56. otes the pixel column 1 lt i lt 64 j the pixel row 1 lt j lt 64 k the frame number 1 lt k lt 32 and m 1 lt m lt M identifies the image sequence The number of image sequences M is determined in the acquisition program values of M gt 40 are recommended for keeping the bias and uncertainty small The subscript s 1 for images recorded with the signal detail in the phantom and s O for images recorded with the detail removed Two and three dimensional discrete Fourier transformed DFT images are denoted with capital letters and the horizontal vertical and temporal frequencies are denoted by u v and w respectively e g G u v F g i j and G u v w F g i j k where F denotes the n dimensional DFT operation It is emphasized that FluoroQuality uses the symmetric normalisation convention of DFT therefore the transformation differs from the DFT with the more commonly used unsymmetrical normalisation convention by the factors V64 64 and 64 64 32 for the 2D and 3D cases respectively Thewidth and height in units of length of the analysed 64x64 pixel sub image is denoted as X and Y respectively However X and Y are not actually measured in FluoroQuality and a numerical value of 1 is used for both of them If the user wishes to express the spatial frequencies and the NPS in proper units mm and mm the values of X and Y can be taken into account by hand calculation for example the horizontal
57. oth the spatial DC frequency and the maximum vertical frequency The noise may often be excessive at the DC channel and the same is often true for the frequency 0 Vax for interlaced imaging systems Including these uncertain channels in calculating the quasi ideal observer s information may impair its performance unnecessarily A similar approach can be used also in other cases where there are isolated frequencies with excessive noise 17 STUK A196 displayed image guality in display devices by using a CCD camera to view the display Chakraborty et al 19993 I n these methods the measurement of MTF NPS or K is not needed but the measurement can be performed simply by applying the DC suppressing observer s detection algorithm to images that are acguired both with and without the detail object in the phantom The DC suppressing observer is constructed by first obtaining e g by averaging of a large number of images approximately noiseless reference images of the phantom in both cases the detail present in the phantom and the detail removed Then by denoting the difference of these averaged images by Ag the decision function is obtained as 1 9 Pres 2 Zasu ku i kl where g denote the pixel values of each image analysed for signal presence and P isthe number of pixels in the image area analysed The SNR is estimated from a set of signal and background images as shown in Eq 1 An advantage of this method is that the resul
58. poral NPS data is not necessary if the user is only interested in SNR measurements The SNR calculation is not performed if the data have already been analysed i e if the AnalysedData folder already contains the file xyzxyz txt the summary file of SNR data The NPS calculation is not done if the AnalysedData folder already contains the files xyzxyzbg spt and xyzxyzsg spt the spatio temporal NPS for the background and signal image series respectively or if the NPS calculation check box is unchecked The calculation can be stopped by clicking the Halt button at any time when the button is enabled The remaining data can be analysed later by choosing the same set of PM files for computation The Retrieve data button is used for examining data that have already been analysed When clicked an open dialog window is shown where the user can choose one of the data sets that have been analysed earlier files indicated by their analysed data file name xyzxyz txt The images in the Analysis form will be updated and the xyzxyz txt file is shown in the memo window The View images button can be used for viewing the original image data series if they havent been deleted from the computer The data can be viewed either as individual frames or as a continuous loop with an user specified frame refresh interval specified in milliseconds The average signal image or background image can be subtracted from the images which may for example help
59. r in Eg 4 represents the expected image signal 6 AGC f K MTF f f JAS f f This expected image signal can be directly obtained also from the difference of averaged background and signal images the directly measured AG can be used instead of both the system transfer characteristics and the signal model in Eq 4 In digital images the ideal observer s SNR can then be obtained as the sum lac 7 SNR eat X o T W where AG is the signal spectrum at spatial frequency f f E and W is the f th component of the noise power spectrum We shall refer the factors SNR 4G W to as the ideal observer s SN R spectrum it shows the contribution of each spatial frequency component to the total SNR jer The practical difficulty of measuring SNR jag with this approach is in obtaining a sufficient number of images for the averaging so that the error from residual noise would be small Gagne and Wagner 1998 n addition to the random variability in the results the residual noise also causes a positive bias to the SN R estimate According to the theory of Gagne and Wagner this bias depends on the number of image samples and the biased estimate of the SNR2 spectrum can be corrected toa de biased estimate by 2N 3 2N 2 8 SNR assi JS ma N where N denotes the number of signal and background image samples in the measurement total number of images is 2 N The bias is slightly different at the zer
60. r than a pixel both of the approximate solutions using the averaged digital MTF or the presampling MTF can result to erroneous conclusions of system performance In measuring the SNR it may not always be necessary to consider the strict ideal observer Other computational observers such as the non prewhitening matched filter NPWMF Wagner and Brown 1985 perceived statistical decision theory model Loo et al 1984 the NPWE model Burgess et al 2001b the Hotelling observer Smith and Barrett 1986 the channelized ideal observer Myers and Barrett 1987 the DC suppressing observer Tapiovaara and Wagner 1993 and the DCsHF s observer who suppresses the information in two isolated spatial frequencies 0 0 and 0 v J Tapiovaara 1997 have been suggested for sub optimal alternatives among others A number of publications e g Loo et al 1984 have shown the close relationship between the performance of such observers and human observers The DC suppressing observer has been used for measuring image quality in fluoroscopy by a PC frame grabber system in laboratory Tapiovaara 1993 and clinical Tapiovaara et al 2000 settings and for image quality measurement in a digital radiography system Gagne et al 2001a A related methodology has also been applied for evaluating phantom images in mammography Chakraborty 1996 and 1997 and the measurement of the This observer can be called the DCsHF s observer because it suppresses b
61. r to trace them to the original publications for the original references and a more thorough and detailed presentation we refer to the scientific literature on the subject A general reference for the concepts of image measurements and image guality is the book of Dainty and Shaw 1974 Barrett and Myers 2003 present a thorough mathematical treatment of the subject Books that consider specifically medical imaging are e g Barrett and Swindell 1981 ICRU 1996 and Beutel et al 2000 Useful general textbooks on statistical decision theory are e g Green and Swets 1966 van Trees 1968 and Whalen 1971 2 1 General concepts of image guality Imaging is basically a process consisting of two distinct stages image recording and image display Wagner 1983 ICRU 1996 This division is especially important in digital imaging where these stages are clearly separate In this case when evaluating image quality one must first decide what one means by the image the acquired image data in the computer memory or a given displayed version of the data Here the word image refers to the acquired image data For discussing image quality one also needs to define quality mages are used for various purposes This suggests that in order to define the concept of image quality in a reasonable manner the underlying task of using the image should be specified an image can be defined to be of good quality if it fulfils its inten
62. rate Eff Doserate 1 mSv 0 40 60 80 100 X ray tube voltage kV Figure 8 The dose to information conversion coefficient for detecting a 13 mm long piece of the USCI 5F catheter in a 25 cm thick patient In this figure the rate of effective dose is used and the x ray projection is heart PA see text for details x total filtration 2 1mm Al p total filtration 2 1 mm Al 0 25 mm Cu curves fitted by hand 42 STUK A196 From the data in figure 7 it is seen that filtration has a large effect on the entrance dose based imaging efficiency whereas the curves are relatively flat as a function of the tube voltage If the skin dose of the patient is of importance as it may be in lengthy interventional procedures it is then best to use a highly filtered x ray beam and an x ray tube voltage between 50 and 70 kV the optimum is at 55 kV At these conditions the entrance dose rate is lowest for the given detectability of the catheter If the skin dose cannot be high enough to cause deterministic radiation effects in the skin it is more reasonable to optimise the imaging conditions by using the effective dose as has been done in figure 8 Again the optimum among the tested alternatives is achieved by using the additional copper filter but now the optimum x ray tube voltage is even lower than it was when the optimum was based on the entrance air kerma rate The effect of filtration on the dose to information conversion coefficient is much
63. re achieving a good repeatability of SNR results may be difficult Therefore it is recommended to use a homogeneous phantom for most measurement applications Unfortunately because of the lack of a standard for recording image data FluoroQuality is not presently a stand alone program Before one can use FluoroOuality to analyse the image sequences one must first record the image data digitally with a specified file format see Appendix D This recording of image sequences must be made by using another program referred to as the acquisition program in this report The acquisition program may record the sequences digitally from the analogue video signal by using a PC frame grabber system or prepare the image data to be analysed from digital image sequence files of digital x ray equipment The FluoroQuality program then calculates and displays the ideal and quasi ideal observers SNR of single STUK A196 frames their SNR e of the acquired image sequences the lag factor and the noise power spectrum NPS of the acquired image data The program also displays the average image for the signal and background situations the net signal and the ideal observer s SNR and SNR e spectra and can be used for visualising the acguired image data STUK A196 2 Background Only a short review of the concepts of image guality that are reguired as a background will be given here It is not attempted to present the evolution of the ideas and concepts o
64. roscopic imaging with image sequences replayed in a continuous loop human efficiency was found to be 30 40 when the display contrast gain was sufficiently high Tapiovaara 1997 I n some other cases for example when comparing images where the observer s efficiency is different e g because of different contrast or noise texture the ranking of the images by a human observer s performance may differ from the ranking predicted by the SNR measurement Human performance is not well understood for many clinically relevant tasks and the relevance of the above objective measurements to human observer performance is not clear in all cases Metz et al 1995 stress that the assessment of medical imaging systems requires going also beyond phantom laboratory measurements into the clinical setting where clinical performance can be assessed by ROC studies for example The same conclusions have been reached in ICRU 1996 It is also noted that even good quality image data can be easily spoiled at the display stage Therefore it is almost a necessity that images are also assessed visually at some stage of the evaluation process There are several ways with which a visual evaluation of image quality can be made with a varying degree of sophistication Presently the Receiver Operating Characteristic ROC and Multiple Alternative Forced Choice MAFC tests are considered to be the best methods of obtaining quantitative and in a less strict sense of the
65. s subjected to a test of the eguality of the variances of the decision variable D with the conditions of signal present and absent the variances should be equal if the noise is truly signal independent The user is noted if one of the variances is more than 20 larger than the other The measurement is also subjected to a X test of the normality of the conditional decision variables The user is warned if the test suggests non normality of the data i e if the value of the X 2 variable is larger than the 1 significance limit N To suppress only the frequency 0 0 it is sufficient to subtract the average pixel value of the entire net signal image However interlaced scanning TV systems often exhibit excessive noise also at the spatial frequency 0 v and it is useful to suppress this frequency as well s 29 STUK A196 3 10 SNR The SNR te of the DCSHF s observer is estimated in two ways i by measuring the SN R of the image sequences essentially as explained in 3 9 but by testing images that are averaged over the whole image sequence length 32 consecutive frames instead of testing the individual frames as was donein 3 9 the SNR e is obtained by dividing the SNR esti mate of the averaged images by the acquisition time of the sequence T The other method ii is to estimate the SNR _ from the SNR in single image frames and the lag as rate R 35 SNR ve posts SNR pcsnrs Lag DCSHFs excl axes 3 11
66. signal detection from several alternatives in a large image area does 38 STUK A196 5 An example of using FluoroQuality for imaging technigue optimisation In this chapter we give an example of how FluoroOuality can be used for optimising the imaging technigue for a given detection task This example is related to interventional imaging and considers the detectability of an USCI 5F catheter type O8LF 0878 in an about 25 cm thick patient The measurement is easy and quick the measurements described here were made in less than two hours The patient simulating phantom consisted of several slabs whose total thicknesses were 20 5 cm polymethyl methacrylate PMMA 4 3 mm aluminium and 5 mm water the lateral extent of the phantom was 24 cm x 24 cm A 13 mm piece of the catheter was filled with water and immersed inthe water layer on top of the phantom in order to mimic its contrast within a vein in this case without any contrast material The distance between the x ray tube focal spot and the image intensifier entrance plane was 110 cm and the field size at the image intensifier entrance plane was 20 cm x 20 cm The distance between the x ray tube focal spot and phantom top was 87 cm The x ray equipment used for the measurements consisted of a Valmet BR 2001 three phase 12 pulse high voltage generator a rotating anode x ray tube Comet DI 10 HS 22 52 150 total filtration without any added filter 2 1mm Al a Philips Imagica 23 cm imag
67. t at scales of the order of the pixel size Dobbins 1995 Therefore the MTF of digital equipment is usually reported in terms of the presampling MTF which does not consider the effect of the discrete sampling on the image In addition to the MTF the other factor needed for describing the signal transfer in the imaging chain is the contrast transfer mage contrast results from the radiation contrast of the detail and the large area transfer characteristics of the imaging system such as the characteristic curve of x ray film The measurement of the large area transfer characteristics sensitometry should be relatively simple but the determination of the radiation contrast of the detail may be difficult it depends eg on the x ray spectrum the attenuation of the radiation in the phantom or patient and in the detail considered the amount of scattered radiation in the image and the photon energy response of the image receptor Image noise is often evaluated visually by determining the threshold contrast Mathematically the image noise of stationary imaging systems can be characterised by the noise power spectrum NPS Wiener spectrum In projection radiography the NPS represents the noise power at various Spatial frequencies specified by f and fy the horizontal and vertical spatial frequencies In fluoroscopy the NPS is three dimensional in addition to the spatial frequency co ordinates one must also specify the noise power as a functi
68. t is not based on any model of observer performance but represents the performance of an actual observer The result may not always be a good estimate of the ideal observer s performance however The ideal observer would outperform it notably if the signal is spread to frequencies where the NPS is strongly frequency dependent This may also be the casein signal dependent non additive noise situations where the ideal observer s strategy differs from the filtering scheme described above The above method Eqs 1 and 9 applies to static x ray images To measure the information relevant to the detail detectability in fluoroscopy one must determine the accumulation rate of SNR SNR xe This quantity is the live image analogy of SNR in static imaging and is required in fluoroscopy because the information obtained depends on the length of the image sequence in fluoroscopy the SNR in a reasonably long image sequence is equal to SNR e multiplied by the imaging time There are at least two approaches to measure this quantity Oneis to record reasonably long fluoroscopic sequences of a duration from one to a few seconds to calculate the time averaged mean images calculate the SNR toa set of such averaged images and finally divide the SNR by the acquisition time The other method involves the measurement of the single frame SNR as for static radiographs above and calculating the N In principle this relationship holds also for short ima
69. ted as 16 W u X v 0 R m uV Again the measurement is made separately for both the signal s 1 and background images s 0 and the value 1 is used for the factors X Y and T in the displayed data Therefore if the user wishes to normalise his her data to the actual size of the image and the temporal length of the sequence the calculated NPS values should be multiplied with the measured value of XYT The spatial frequencies are obtained by dividing the integer values of u and v by X and Y respectively 24 STUK A196 3 5 The full spatio temporal NPS The full spatio temporal NPS is calculated as XYT W u X v Y w T 17 3p U v w M 1 64 32 Elz leli jak m g JN Again the measurement is made separately for both the signal s 1 and background images s O and the value 1 is used for the factors X Y andT If theuser wishes to normalise his her data tothe actual size and temporal length of the image sequence the calculated NPS values should be multiplied with the measured value of XYT and the spatial frequencies are obtained by dividing the integer values of u and v by X and Y respectively In displaying the NPS the temporal length of the image sequences is already taken into account and the proper temporal frequency in Hz is displayed on the NPS display form 3 6 The SNR measures obtained by integration analytical biased data Using the measured signal spectrum and NPS
70. the SNR with visibility Actually there is no detectability threshold but the detail turns from not visible to clearly visible through a continuosly improving detection certainty when the SNR is increased Thestatistical efficiency F d SNR jaa of humans is often found to be of the order of 50 then a human observer who is presented images with a detail of say SNR jay 2 would obtain a d 1 41 and achieve a 84 Yo probability of a correct answer in a 2AFC test or a 39 probability in a 16 AFC test see figure 5 for the relationship between d and the probability of a correct response in some MAFC tests Some authors have equated the criterion of 1 2 agi PA L a a 7 7 7 0 8 YAWA oO 2 F 4 O gt 0 6 JA g 128AFC O LA 2 0 4 TBIDAEC o pa O L E YA o 0 2 AA L 7 ky A Pd A o pee 0 2 4 6 d Figure 5 The relationship between the observer s signal to noise ratio at the decision level d and the probability of a correct response in some MAFC tests The use of SNR thresholds has been criticized by many researchers e g Burgess 1983 who also pointed out that Rose had suggested SNR threshold values denoted by k in the range from 3 to 5 After his criticism Burgess writes H owever if you insist on using the Rose model it is suggested that you use values of k in the range from 5 to 10 for simple detection and 15 to 20 for signal identification t
71. the actual location of the point with respect to the pixel boundaries One possible solution for measuring the MTF in such a system is to locate the stimulus at all possible locations within the pixel boundaries this can be done using e g a slightly angulated slit and calculate the average of these different MTFs However this average digital MTF then no longer is related to the point spread function at any location and strictly cannot be used for comparing the sharpness of two systems Another possibility is to measure the presampling MTF Fujita et al 1989 Dobbins 1995 concludes that in the common case of undersampled digital imaging the interpretation of NEQ is difficult and depends on the measurement method and the frequency content of the incident information He suggests the use of the averaged digital MTF for calculating the NEQ In other publications variable definitions of NEQ have been used but the use of the presampling MTF seems to become the most common convention However there iS no unambiguous solution for interpreting the NEQ or DQE results at frequencies where aliasing effects are important 15 STUK A196 The above problem can be avoided in measuring the SNR if the SNR measurement is done more directly not by going through the transfer function analysis but by measuring the detectability of the detail as based on how the detail is actually imaged the problem is circumvented It is noted that the nominato
72. ts the files xyzxyzPM dat xyzxyzYY ims and xyzxyzY Y imt must be kept YY is 01 M where M is the number of image sequences The contents of datafiles e Average image data are saved as AnalysedDatal xyzxyzsg ave and AnalysedData xyzxyzbg ave for the signal and background sets respectively The files contain the pixel values as eight byte real numbers line by line starting from the left upper corner Thetemporal O freguency NPS of the signal and background image sets are saved as AnalysedData xyzxyzsg nps and AnalysedData xyzxyzbg nps The files contain the NPS values as eight byte real numbers line by line starting from the 0 0 frequency e Theone frameNPS of thesignal and background image sets are saved as AnalysedData xyzxyzsg npl and AnalysedData xyzxyzbg np1 The files contain the NPS values as eight byte real numbers line by line starting from the 0 0 frequency The origin of Fourier transformed data that are displayed on the form is in the middle of each image In the data files the origin can be thought of as being located in the upper left corner Note the periodicity of theDFT data 59 STUK A196 WHICH FILES ARE NECESSARY TO KEEP AND APPENDIX C SNR measurement data is saved in the text file AnalysedData xyzxyz txt e The SNR data displayed above the memo box in the basic form are contained in the text file AnalysedData xyzxyzSN Rdataonform dat The spatio temporal
73. uences of 32 consecutive frames each were 30 STUK A196 recorded for both the signal and the background case The true value of the SNR for each disk at each separate imaging condition is obtained by physical reasoning as the average of the measurements by methods ii and iii above for thethickest detail at that imaging condition and scaling this SN R by tothe ratio of the squares of the disk thicknesses The SN R2 estimates are shown in Fig 1 and the SNR estimates in Fig 2 rate 1000 x w 100 E n p tr a a 10 a Biased PWMF estimate Ch 3 6 Wt A De biased PWMF estimate Ch 3 7 x gt x Template method Ch 3 9 1 10 100 1000 SNR Figure 1 The estimates of the SNR in single image frames as measured by three methods Chapters 3 6 3 7 and 3 9 The data correspond to the detectability of 1 cm diameter PMMA disks of various thicknesses in various imaging conditions Number of signal and background image files M 40 The data are plotted against the true SNR of the details 31 STUK A196 10000 w 2 1000 av E n 8 Biased PWMF estimate Ch 3 6 100 Tr De biased PWMF estimate Ch 3 7 a Template method Ch 3 10 10 10 100 1000 10000 SNR ate 1 5 Figure 2 The estimates of the SNR as measured by three methods Chapters 3 6 3 7 and 3 10 The data correspond to the detectability of 1 cm diameter PMMA disks of various thi
74. ve as much as would be expected by SNR reasoning The threshold SNR va variability between the observers in these tests was large about 40 and one should consider these thresholds only as typical values Nevertheless these figures may give a feel for the SNR e magnitude required for a subjective sensation of visibility We also made 16 AFC tests at two of the lowest dose levels referred to above The SNR xe of the test detail was equal to 19 s at the lowest dose rate and equal to 40 s at the two fold dose rate level i e the SNR x of the test details were about one third or one half of the SNR x of the details that the observers reported as just visible in the threshold experiment The square of the observers mean detectability index d was 6 6 and 12 4 in these imaging conditions respectively These results are compatible with the assumption that the observers d is proportional to the SNR j This suggests that in the threshold test the observers evaluated the detail contrast to be more important than was actually the casein the controlled detectability test The observers were instructed to try to see through the noise but without needing to guess 37 STUK A196 It is problematic to evaluate the statistical efficiency of the human observers in fluoroscopy because the available SNR increases with fluoroscopy time and the efficiency would be very low but one can think of theratiot d 7 S
75. weaker for the effective dose rate based evaluation than it is for the entrance air kerma rate case Not much advantage is gained by adding filtration in fact the imaging efficiency of the more heavily filtered radiation appears to be lower than for the lightly filtered radiation when the tube voltage is high The imaging efficiency is seen to depend strongly on the x ray tube voltage the dose rate reguired for a constant detectability of the catheter is much higher at a high x ray tube voltage than it is at voltages near 50 kV The above discussion considers only the imaging efficiency SNR dose rate For an actual optimisation of imaging conditions one needs also to specify the image quality that is required in the dinical procedure and take into account the constraints imposed by the x ray system For minimising the dose it may be even more important to keep the image quality as low as is sufficient than to work at the exactly optimum imaging efficiency conditions of course however the quality must be high enough to ensure the proper performing of the examination or procedure If good quality imaging is necessary the thermal load of the x ray tube may become too high at the technique using heavy filtration and a low tube voltage or the fluoroscopic current may be limited such that the required image quality cannot be achieved at these imaging conditions Then one must use less filtration or a higher tube voltage than the optimum effici
76. which this quality can be obtained I n choosing the appropriate image quality level one must then weigh the potential risk from the loss of diagnostic information in the low dose application against the larger radiation risk from higher dose techniques Martin et al 1999 19 STUK A196 2 4 Visual assessment Metz et al 1995 have reviewed the assessment of medical image guality and noted that there exists a wide consensus in measuring the sensitometric guantities MTF and NPS of radiological systems They also agreed that the combined measures NEQ DQE and SNR joa the ideal observer s signal to noise ratio are useful for normalising the measurements on an absolute scale and for relating those measurements to the decision performance of the ideal observer However they stress that in the two stage recording and display description of the imaging process SNR ja describes image quality at the stage of image recording This can be considered an advantage for understanding the steps through which images are formed but the data stage cannot be used aloneto predict the ranking of images that a human might make on basis of the displayed image if the characteristics between the images are too different In many cases however such as projection radiography the human and ideal observer results show a good correlation and it seems that the efficiency of human observers is of the order of 50 Similar observations have been made also in fluo
77. word objective results of human observers ability to detect signals in the images The results from these tests can be given in terms of the decision stage SNR of human observers which is often denoted ideal 20 STUK A196 as d These psychophysical methods can be used for both clinical studies of actual patient images and detection tests using simple phantom radiographs but they are not suitable for e g routine quality assurance work Therefore more simple but less accurate methods need often be used eg subjective assessment of detail detectability in phantom images Often these phantoms and details are highly simplified and the detection task may not be reasonably related to clinically meaningful tasks Typical examples of common image quality measurement tools are line pair test plates and contrast detail phantoms whose images are visually evaluated For a more detailed discussion on visual evaluation methods see for example ICRU Report 54 and the references therein 21 STUK A196 3 The image quality quantities measured with FluoroQuality 3 1 Notation and conventions FluoroQuality analyses only a part of the whole image area The analysed area Sub image is selected in the acquisition program by the user These sub images must be of size 64x64 pixels with 8 bit pixel depth and each recorded sequence must contain 32 consecutive image frames These image data are denoted bel ow by g i j k m wherei den
78. y F Verdun FR Hessler C and Schnyder P Estimation of the noisy component of anatomical backgrounds M ed Phys 26 1365 1370 1999 Brown DG Insana MF and Tapiovaara M Detection performance of the ideal decision function and its McLaurin expansion Signal position unknown J AAcoust Soc Am 97 379 398 1995 Burgess AE Observer performance testing for medical imaging Notes for A A P M refresher course August 1983 Burgess AE Comparison of receiver operating characteristic and forced choice observer performance methods Med Phys 22 643 655 1995 Burgess AE Wagner RF J ennings RJ and Barlow HB Efficiency of human visual signal discrimination Science 214 93 94 1981 Burgess AE Jacobson FL and Judy PF Lesion detection in digital mammograms Proc SPIE 4320 555 560 2001a Burgess AE J acobson FL and J udy PF Human observer detection experiments with mammograms and power law noise Med Phys 28 419 437 2001b Cahn RN Cederstr m B Danielsson M Hall A Lundqvist M Nygren D Detective quantum efficiency dependence on x ray energy weighting in mammography Med Phys 26 2680 2683 1999 Chakraborty DP Physical measures of image quality in mammography Proc SPIE 2708 179 185 1996 46 STUK A196 Chakraborty DP Computer analysis of mammography phantom images CAMPI An application to the measurement of microcalcification image guality of directly acguired digital images M ed Phys 24 1269 1277 1997a
79. y be high The SNR ja for a specified SKE BKE detection task described by the difference of the signal and background inputs As x y can be calculated as 3 DQE f fy 2 K MTF SG oF ASC fy Wief 4 SNRiiea ff df df 14 STUK A196 where AS f_ f is the Fourier transform of As x y One can also express the DQE as a task related quantity as DQE JIN SNR vical image a N SNR eat in where SNR jat image ANA SN R jaa in relate to the detectability of a given detail as based on the image data and the radiation incident on the image receptor respectively Tapiovaara and Wagner 1985 The above eguations 2 4 have been written here for the case of analogue images A treatment involving the discrete pixels of digital imaging modalities has been used for example in ICRU report 54 Myers et al 1987 and Tapiovaara and Wagner 1993 n this treatment images are represented by vectors whose dimensionality corresponds to the number of pixels in the image Digital imaging poses some problems for NEO and DOE measurements Dobbins 1995 Pineda and Barrett 2001 Gagne et al 2001a and b The main problem is undersampling which when present resultstoaliasing TheMTF is then no longer a transfer factor of a given frequency Aliasing can also be described as the conseguence of the violation of the assumption of shift invariance which would be reguired in theMTF analysis theimage of a point may depend on
80. y not be large enough toremove the bias of very faint details One possibility to measure 34 STUK A196 the SNR of such thin details might be to derive their SNR from the measurement of a thicker detail by scaling with the ratio of the squares of the detail thicknesses if possible Overall the conclusions of the SNR e estimates are similar to those above the bias is minor for reasonably strong signal details and an important issue only for very faint signal details The analytical de biased estimates are slightly positively biased the templatemethod estimates are slightly negatively biased and the bias decreases with an increasing number of image samples In this case however the bias of the template matching estimate appears to be lowest for all the details when M is low Based on the above results and our earlier experience the DCSHFs template based measurement method seems to result in the most reliable SNR estimates from the measurement alternatives considered here The use of a low M and a very weak signal detail should be avoided The bias which is evident at low SNR and SNR xe Values for moderate M could perhaps be reduced by using the average of the template based and the analytical de biased estimates 35 STUK A196 4 SNR and detail visibility For static radiographs it is often said that a SNR of 3 7 SNR 9 50 is needed for a detail to be visible This is however just a rule of thumb trying to relate
81. y value by the width or height of the analysed image area in mm respectively Clicking inside an image results to a zoomed copy of that image in the lower left image box 55 STUK A196 FLUOROQUALITY v 2 0 User s Guide APPENDIX A e Double dicking in the Memo results to a copy of the memo text to be copied on the Clipboard 5 The SNR SNR x and lag measured by the template method are indicated on the left side of the basic form above the memo Contents of the images on the analysis form The contents of the image data are explained by their labels The displayed images are The sample averages of all signal and background images in terms of pixel values Thetemporal 0 frequency NPS of the signal and background image sets The values arein terms of pixel values and the normalisation does not include the spatial and temporal extent of the images Nominally a1 mm area and 1 s duration is assumed to normalise properly to units of mm s the data should be multiplied by the area of the images in mm and the temporal length in s Theoneframe NPS i e spatio temporal NPS summed over all temporal freguencies of the signal and background image sets The values are in terms of pixel values and the normalisation does not include the spatial extent of the images Nominally a 1 mm area is assumed to normalise properly to units of mm the data should be multiplied by the area of the images in mm The net signal

STUK - Julkari

Contents

Download Pdf Manuals

Related Search

Related Contents