Development of “filmless” cardiac catheterization laboratories is eminent. The problems of implementing a digital catheterization laboratory involve archiving large amounts of data per procedure and high transfer rates to retrieve previous procedures. Lossy compression can accommodate these changes, but at the cost of possibly impairing detection of clinically important angiographic features. Our study involves the observer detection and classification of features in clinical images and the effects that JPEG and wavelet compression have on the detectability of these features. We found no significant degradation in human observer performance with 7:1 and 15:1 JPEG compressed images in 6 clinically relevant visual tasks. Human observer performance for wavelet compression degraded significantly for 2 out of 6 tasks at 7:1 compression and 4 out of 6 tasks at 19:1 compression.
©1999 Optical Society of America
The use of digital angiographic cardiac catheterization laboratories in clinical practice has readily been embraced by cardiologists. Digital images allow improved image quality by utilizing high-pass filtering techniques to sharpen images, contrast enhancement techniques to maintain proper differentiation of neighboring anatomical areas under varying exposure conditions, and variable rate playback of image sequences to improve feature detection. With the acceptance of all digital Picture Archiving and Communications Systems (PACS) in Radiology, the advent of Cardiology PACS systems is inevitable. Some of the problems encountered in PACS systems are the large archival storage requirements and the need for expedient retrieval of patient images. Long term storage requirements are challenging because of the number of patients performed in one year. A typical four procedure room catheterization laboratory can perform approximately 3000–4000 procedures a year. In addition, the shelf life of image sequences are long for children’s catheterization procedures which must be stored until they reach 21 years of age. Adult cine angiograms must be stored for 7 years after the procedure. Usually storage of old cine angiograms is off-site from the hospital and can take on average one day to retrieve them. This is not a problem for digital off-site storage of the image data, because you can retrieve the data over the wide area network. Without the use of compression, the transfer rate can be exceedingly slow. Another benefit of digital storage is the availability of the images to multiple physicians. Many times a cine film is checked out or lost and therefore not available to other clinicians. The use of digital storage allows multiple copies of the original data to be available. The large amount of digital data required for an individual cardiac catheterization procedure (~1 Gigabyte) and the high data rates required for real-time display (512×512×8 bits at 30 frames/sec) can easily overwhelm a PACS cardiology system . Image compression algorithms provide a means for reducing storage and communication requirements. This reduction in storage and communication throughput is not without a cavet. Lossy image compression can degrade diagnostic accuracy depending on the amount of compression. Measurements of root mean square error of the uncompressed and compressed images does not agree well with human perception studies . The difficulty is in measuring the degradation in diagnostic accuracy for a particular task.
There are only a few studies on the impact of lossy image compression on diagnosis of various morphological features in coronary angiograms [3–5]. Rigolin et al  found that stenosis (vessel narrowing) severity and quantitative coronary diameter measurement precision did not degrade with 15:1 JPEG compression versus uncompressed images. Baker et al  focused on more subtle morphological features such as dissections, thrombi, and coronary stents and found no significant degradation from the original to 15:1 JPEG compressed angiographic images. The conclusion from these studies is that 15:1 JPEG compression may be adequate for clinical decision making. In both studies a panel of “expert” observers determined the gold-standard truth in the observed images. Images where no consensus was achieved were excluded from the study. It is therefore plausible that many of the images with subtle lesions would result in no agreement among committee members and therefore excluded from the study. In addition, a recent study using simulated signals embedded in x-ray coronary angiograms has shown that consensus by an expert committee can lead to a false gold standard and therefore might lead to erroneous conclusions about image quality evaluation .
The purpose of the present study is to use computer-simulated lesions embedded in real x-ray coronary angiograms to evaluate the effect of image compression on feature detectability and classification. This latter method does not present the problem of potential false gold standards established by consensus committee. Although, it might be argued that the simulated lesions approach is not as realistic as using real lesions. Our approach hypothesizes that the visual and cognitive processes underlying the detection of the computer simulated lesion are the same as those mediating the real clinical diagnosis. Any detriment in the detectability of fundamental features will also degrade diagnostic decision. We have shown that image display optimization based on experiments with simulated lesions generalize to actual clinical tasks .
In addition, the previous studies that evaluated the affect of image compression used a fixed JPEG quality level or factor to compress all image sequences in the study [3-5]. A fixed quality level will result in different compression ratios for different image sequences because of the variation in image content. For example, compression of 100 images with a quality level of 50 in our laboratory resulted in a compression ratio of 20.4±4.2 (with a bimodal distribution far from Gaussian). In the present study, each image was iteratively compressed at different quality levels until the desired compression ratio was achieved. Therefore, the present study has less variability on the achieved JPEG compression ratios making our evaluation more precise than previous ones. Finally, the present study also extends previous work on JPEG by evaluating the effect of wavelet image compression not studied previously for x-ray coronary angiograms.
The simulated arteries and lesions are modeled after x-ray coronary angiograms. Our simulated artery model is based on three major components. First, we model the attenuation of the x-rays as a function of path length through the contrast filled vessel. Second, we model the image system blur (focal spot and image intensifier distortion) as a spatial convolution of the simulated artery with a 2-dimensional isotropic Gaussian point spread function. Finally, we estimate the scattering and veiling glare and correct the image based on the convolution technique of Love and Kruger . Individual details about the simulated artery model are discussed in the Appendix.
The morphological features or lesions in our individual studies were bridging lumens, stenosis, thrombosis, and ulcerated plaques. An individual image was divided into 4 quadrants. The four simulated arteries were placed into one of the quadrants. For our test-images, the projected simulated arteries are 3-dimensional right circular cylinders with a normal diameter of 12 pixels (3.6 mm). The stenosis had an hourglass shape and was generated as a sinusoidal modulated tapering of the vessel diameter with a minimum diameter of 8 pixels. The overall length of the simulated artery was 50 pixels (15.0 mm). The attenuation coefficient, µ, was set to 0.16/mm to produce simulated arteries with the same projected intensity as clinical angiograms of coronary arteries of the same diameter. The simulated arterial segments were generated at four times the normal size and then down-sampled. The final overall image size presented to the observers was 512×512×8 bits in dimension.
Figure 1 shows a dynamic demonstration of the generation of test images. It includes the following: projection of simulated normal and abnormal arteries (with ulcer) at four times the resolution, focal spot/image receptor blur, sub-sampling of simulated artery, estimation of secondary and primary components of an x-ray clinical background, and final placement of the artery into the clinical background with the proper scatter and veiling glare estimate added back into the image.
2.2 Clinically relevant tasks
There were four different clinically relevant tasks in the study: lumen, graded stenosis, thrombosis, and ulcer. Of the four tasks, two tasks (graded stenosis and thrombosis) were evaluated for static (single frame) viewing and dynamic (32 frames at 16 frames/sec) viewing. Table 1 lists the individual tasks and display conditions (static vs. dynamic) evaluated in the present study.
2.2.1 Detection of morphological features
There were three different morphologic detection tasks. For the three tasks, on each trial four simulated arterial segments were presented. The observers’ task was to correctly select the arterial segment (one out of four arteries, 4 alternative forced choice) containing the morphological feature (Fig. 1). The different tasks were based on the morphological feature to be detected. All features were inserted into the simulated arteries by considering 3-D models of the lesions and calculating the change in the incident absorption due to the presence of the lesion.
2.2.2 Thrombus task
The thrombus represents coagulated blood in the vessel. A filling defect within the artery does not allow the contrast agent to full opacify the artery and therefore results in lower absorption of the x-ray beam. The filling defect appears lighter within the dark contrast filled vessel. The thrombus detection target was represented as a simulated filling defect with a hemi-ellipsoidal shape located at the vertical and horizontal center of one of the four simulated arteries and assumed to have a diameter of 6 pixels. Therefore, the projection of the thrombus appears as a brighter spot within the artery (Fig. 2A).
2.2.3 Ulcer task
The ulcer represents an aneurysm; an outer-pouch attached to the artery that’s filled with contrast agent. The ulcer detection target (Fig. 3A) was a sphere with a diameter of 8.0 pixels attached to one of the simulated arteries. The ulcer had an overlap of 4.0 pixels with the artery on the imaging plane and was always located 15 pixels away (vertically) from the top of the artery. We consider an ulcer that protrudes from the artery and is located perpendicular to the x-ray beam. This ulcer would then appear as a semi—circular attachment to the border of the artery (see arrow in Fig. 3A).
2.2.4 Bridging lumen task
The word lumen refers to a portion of a faintly visible artery. The low contrast location within the artery is due to blockage or narrowing of the artery by blood clotting, atherosclerotic plaque, or lesion formation. The lesion prevents the contrast agent from filling the full diameter of the artery and therefore reduces the attenuation and contrast of the projected artery. The word “bridging” refers to the fact that the low contrast-diseased portion of the artery generally connects or bridges to high contrast portions of the artery.
In our bridging lumen detection task, the simulated arteries had a minimum diameter of 6 pixels. Three of the four simulated arteries had a 5-pixel gap at the point of minimum diameter. The remaining artery was intact and therefore had a low contrast lumen bridging the 5-pixel gap. The observers’ task was to select the artery with the bridging lumen (see Fig. 4A).
2.2.5 Classification of stenosis grading
The stenosis grading task (Fig. 5A) was different from the previous task because the observers had to classify the narrowing of the artery (stenosis) into one of four categories. As in the previous detection tasks, each image we embedded with four simulated arterial segments. In this task however, the minimum diameter of each artery was randomly selected (with uniform probability) to be from 0 to 100% of the normal portion of the artery (12 pixels). The vertical length of the sinusoidal narrowing was 24 pixels. The task of the observers was to classify each of the four arteries in the image in one of four categories: normal (0–25 % stenosis), mild (25–50 % stenosis), moderate (50–75 % stenosis), severe (75–100 %). At all times during the experiment a reference image was presented to the observers showing simulated arteries at the boundary of the categories together with category labels (0 %, 25 %, 50 % and 75 % stenosis arteries).
2.3 Image compression
Images were compressed and decompressed with the fifth public release of the Independent JPEG Group’s free software . The wavelet algorithm is a beta release of a software packaged named compression with reversible embedded wavelets (CREW) code developed by the RICOH corporation for continuous-tone image compression of medical images (RICOH Company Ltd., Menlo Park, CA) . The CREW wavelet transforms uses a “Two-Ten” (TT) reversible transform based on a decomposition of the LeGall-Tabatabai polynomial. The TT transform is a specific reversible wavelet pair with 2 low pass and 10 high pass analysis filters. The TT-Transform is reversible, and can be used for lossless and lossy compression. Another advantage of the TT transform is that one of the 10-tap low pass synthesis filters makes the TT-Transform an overlapped transform. The overlapping wavelets eliminate the “blocky” artifact seen in highly compressed JPEG images because of the smoothing of sharp image transitions. Although there are many algorithms available, we chose versions of JPEG and wavelet that were easily available and well documented. The JPEG and wavelet executable-codes were used in their default configurations and were not specifically optimized for medical imagery. For each test-image the quality factor (which ranges from 0 to 100) was iteratively changed until the desired compression ratio was achieved. The target compression ratios were 7:1 and 15:1. The achieved compression ratios were 6.9:1±0.17 and 14.9:1±0.16 for the JPEG algorithm and 6.6:1±0.79 and 19.0:1±4.2 for the wavelet algorithm. The inability to achieve a 15:1 compression ratio and the larger variability of compression ratios for the wavelet algorithm was due to an inherent discontinuity in the quality factor near the 15:1 compression ratio. An example of the uncompressed, 14.9:1 JPEG, 19.2:1 Wavelet are shown in figures 2–5 images B and C. for the four lesion conditions: thrombosis, ulcer, bridging lumen, and stenosis.
2.4 Psychophysical studies
Four non-physician trained observers (PV, CH, GA and GR) and two physicians (DV and JH) with previous experience reviewing x-ray coronary angiograms participated in the studies. The specific distribution of observers for each task is given in Table 2 and Fig. 6. All tasks had three observers except for the grading stenosis in the dynamic images which included only two observers. Data for different tasks were collected separately. Each reading session by our observers involved 100 randomly chosen images from the 424 images for the static images’ study and 100 images for the dynamic display study. For each task observers first trained in 1 session of 100 trials for each compression condition and then participated in 5 sessions of 100 trials per compression condition. Conditions were randomized within a day. Images were displayed on an Image Systems M17LMAX monochrome monitor with a maximum resolution of 1664×1280 pixels (Image Systems, Minnetonka, MN). The mean luminance was 16.0 cd/m2. The luminance vs. gray level relationship was the default nonlinear that would be used by the physicians in a clinic with this monitor. Observers viewed the images binocularly from a distance of 50 cm and had unlimited time to reach a decision. When a decision was reached the observers pressed the number 1, 2, 3 or 4 in the keyboard to indicate their choice for that trial. For the morphological feature detection tasks, the four different keyboard keys corresponded to the four different arterial segments. For the stenosis grading task each of the four keys corresponded to each of the stenosis categories: 1) normal, 2) mild, 3) moderate, and 4) severe. For the dynamic display, sequences were displayed at 16 frames per second.
2.5 Statistical Analysis
Data for each task was analyzed separately. The proportion of trials (Pc) that that observer correctly selected the feature location or classified the arterial segment (for the stenosis grading task) was computed for each session, compression condition and observer. Pc scores were then transformed into an index of detectability (da) as given by Signal Detection Theory  for the morphological feature detection tasks. For visual detection, the index of detectability da, has been shown to be approximately independent of the number of alternatives and linear with signal contrast [7,12]. For the stenosis grading task, conversion to da, with the standard M-AFC equation is not applicable because the stenosis grading is a classification task. Conversion to an index of detectability would involve estimating three detectability indices specifying the distance in standard deviation units between the observers’ internal response distribution to each of the stenosis grading categories. This procedure is more complex and was not attempted .
A single-factor within subjects analysis of variance was performed to evaluate the overall effect of image compression across observers. In the usual fixed effects analysis of variance model, assumptions are made that the observations are independent and normally distributed, and that the variance is the same in all cells. Paired-comparisons between compression ratios were done using a Bonferroni multiple t-test. Three observers were used in each morphological condition (Table 2.). BMDP statistical software version 7.0 was used to perform the previously described statistical analysis on the observer performance data (SPSS Inc., Chicago, IL.).
Figure 6 are graphs of the individual observer performance for each of the observer tasks. Each graph has three observers which are identified by one of three colors: red, blue or green. The solid symbols represent the observer performance reading JPEG compressed images. The open symbols represent the observer performance reading wavelet compressed images. The y-error bars in each graph represent the standard error of the measure. The x-error bars represent the standard deviation of the compression ratio. The overall pattern of results across all tasks and display conditions (static and dynamic) for each individual observer is that the solid lines (JPEG) show a superior performance level when compared to the broken lines (Wavelet).
Table 2 shows the results of the statistical comparisons for all tasks and compression levels for the JPEG and wavelet algorithms. For JPEG compression, we found no statistical significance for any of the tasks between the mean observer performance at 1:1 when compared to mean observer performance at 7:1 and 15:1 respectively.
The overall effect of wavelet compression on observer performance was statistically significant (α=0.05) for all but graded stenosis. Static graded stenosis was not statistically significant at 1:1 versus 19:1 (p=0.063). Bonferroni pairwise comparisons were performed on five of the six detection tasks (lumen, static and dynamic thrombosis, static and dynamic graded stenosis, and ulcer) at 1:1 versus 7:1 and 1:1 versus 19:1. The 1:1 versus 7:1 wavelet compression reached statistical significance for two out of six tasks (static thrombosis and ulcer). The 1:1 versus 19:1 wavelet compression was statistically significant for four out of six tasks (ulcer, dynamic graded stenosis, static and dynamic thrombosis).
In the evaluation of lossy image compression, it is vital that the diagnostic task be representative of that used in the clinical procedure. We used a battery of clinically relevant visual tasks including new morphological features not tested previously (i.e. bridging lumens and ulcerated plaques). Our methodology with the use of simulated lesions does not supplant the methodology of Rigolin and Baker but complements their work. Both approaches resulted in agreement that 15:1 JPEG compression showed no statistical significance for our individual tasks.
The observer performance of our clinical observers JH and DV (Fig. 6) out performed the non-clinical readers in their respective tasks. This indicates that their training as physicians aided in the detection of the lesions.
Based on p-values for JPEG compression, the static thrombosis detection task was the most sensitive to all compression levels (p=0.051). While, dynamic 16 frames/sec thrombosis detection was the least sensitive to all compression levels (p=0.408). For wavelet compression, the results show a larger effect of compression for static thrombus detection than for dynamic thrombus detection. This suggests an interaction effect between the compression and motion of the thrombosis; however, further investigation is needed.
For JPEG compression, three out six of the detection tasks were close to statistical significance. Wavelet compression had one detection task that was close to statistical significance. Perhaps with an increased number of observers statistical significance might have been achieved.
Our results suggest that for all the tasks tested wavelet image compression degrades diagnostic accuracy more than JPEG compression. One of the shortcomings of our study was the inability to match the average compression ratios for the wavelet (19:1) and JPEG (14.9:1) compression ratios at the 15:1 level. This was due to the wavelet algorithm that had a discontinuity in the quality level vs. compression ratio relationship. This is an area of further research to optimize the image quality based on the wavelet parameters such as number of decomposition levels and quantization of wavelet coefficients within each sub-band.
In conclusion, we have extended previous work on the accuracy of visual diagnostic evaluation of JPEG image compressed x-ray coronary angiograms. We presented a paradigm that does not present the shortcomings of a consensus expert committee and a method to achieve higher precision in compression ratios. We have shown the affects JPEG and wavelet image compression have on performance detection and classifying simulated morphological features embedded in coronary angiograms and have found no loss in diagnostic performance up to 15:1 JPEG compression. Wavelet compression was statistically significant under all conditions except for bridging lumen detection at 19:1 image compression. Further work will concentrate on optimizing the quantization tables of both JPEG and wavelet compressions based on observer performance of various lesion detection tasks to increase the compression ratio and maintain image quality. In addition, the use of other types of wavelets such as those developed by Daubechies and Chui might improve observer performance for wavelet compression [14,15].
6. Appendix - Generation of test images
The radiographic imaging chain consists of the following: X-ray generator, object of interest, image intensifier, and digitizer. We developed our test-image algorithm based on the parallel beam X-ray to ease implementation. An analytic study of the effect a cone beam X-ray projection through an artery parallel to the image intesifier (for the worst possible geometry, 9” image intensifier mode, minimum X-ray source to intensifier distance 85 cm, minimum X-ray source to table top distance 45 cm) reveals that the maximum error in the path-length through the artery compared to a parallel beam X-ray projection is less than 1% of the projected diameter. The model decomposes the resulting image I(x,y) into a primary component due to the primary beam, P(x,y), and a secondary due to scattered rays, S(x,y).
The primary beam consists of the X-ray beam passing through the cardiac area and attenuated by bone (sternum ribs and spine), tissue (lung, myocardium and great vessels) and contrast agent (iodine). The resulting primary component of the image is given by exponential attenuation of the incident beam Io (analogous to the X-ray absorption process) by the densities of the projected patient (Dp) and contrast filled artery (Da).
Da(x,y)=m ta(x,y) where m is the attenuation coefficient of the contrast agent (iodine) and ta(x,y) is the artery thickness. Also the patient density is given by the summation across the thickness and attenuation coefficients of the different tissues (bone, iodine and myocardium tissue) transversed by the x-ray beam
We wish to insert a simulated object in the scene so that the primary component of the image needs to be attenuated by the density of the simulated object ():
where Ps(x,y) is the primary including the simulated object. The density of the simulated object is given by Ds=µts(x,y) where µ is the attenuation coefficient and ts(x,y) is the thickness of the contrast filled simulated object transversed by the x-rays as a function of the image position. Since we do not have access to the primary but rather an image with the primary and secondary, we first subtract the secondary from the image, then attenuate the estimated primary due to the simulated artery and then add back the secondary component:
where Is(x,y) is the image after the simulated object is inserted into the image and I(x,y) is the original image prior to insertion of the simulated object. The secondary component S(x,y) is not known apriori but can be approximated based on previous work. Love and Kruger  used imaged phantoms to empirically show that the secondary image component due to scattering and veiling glare can be approximated by the convolution of the image and a double exponential kernel :
where α=0.483, β=7.69. The convolution kernel is given by:
Under clinical conditions we must consider the distortions of the imaging chain. We model the blurring that is introduced at various stages in the imaging process as a convolution of the vessel intensity distribution I(x,y). Focal spot and image intensifier blur are accounted for by convolving the projected cylinders with Gaussian point spread function with standard deviation σ=1 pixel (0.3 mm). We degrade the vessel profile with white noise w(x,y) in order to remove the staircase or digitization effect of the artery profile and to simulate quantum mottle.
This work has been funded by the National Institute of Health (Heart Lung and Blood) ROI HL 53455.
References and links
1. S. E. Nissen, C. J. Pepine, T. M. Bashore, P. C . Block, L. I. Bonchek, J. A. Brinker, B. Carabello, J. S. Douglas, J. L. Elion, J. W. Hirshfeld, D. R. Holmes, W. L. Johnson, W. P. Klinke, D. C. Levin, G. B. J Mancini, C. E Mullins, J. D. Thomas, E. J. Topol, J. H. K. Vogel, and M. A. Wondrow, “Cardiac Angiography Without Cine Film: Erecting a “Tower of Babel” in the Cardiac Catheterization Laboratory,” J. of Am. Col. of Card. 24, 834–837 (1994). [CrossRef]
2. G. Strang and T. Nguyen, Wavelets and Filter Banks, (Wellesley-Cambridge Press, Wellesley, MA, 1996) 362–364.
3. V. H. Rigolin, P. A. Robiolio, L. A. Spero, B. P. Harrawood, K. G. Morris, D. F. Fortin, W. A. Baker, T. M. Bashore, and J. T. Cusma, “Compression of Digital Coronary Angiograms Does Not Affect Visual or Quantitative Assessment of Coronary Artery Stenosis Severity,” Am. J. of Card. 78, 131–135 (1996). [CrossRef]
4. W.A. Baker, S. E. Hearne, L. A. Spero, K. G. Morris, R. A. Harrington, M. H. Sketch Jr., V. S. Behar, Y. Kong, R. H. Peter, T. M. Bashore, J. K. Harrison, and J. T. Cusma, “Lossy (15:1) JPEG Compression of Digital Coronary Angiograms Does Not Limit Detection of Subtle Morphological Features,” Circ. 96, 1157–1164 (1997). [CrossRef]
5. J. S. Whiting, M. P. Eckstein, C. A. D. Honig, S. Gu, S. Einav, and N. L. Eigler, “Effect of lossy image compression on observer performance in dynamically displayed digital coronary angiograms,” Circ. 86 (Suppl.) I-444 (1993).
6. M. P. Eckstein, T. D. Wickens, G. Aharanov, G. Ruan, C. Morioka, and J. S. Whiting, “Quantifying the limitations of the use of consensus expert committees in ROC studies,” Proceedings of the SPIE Medical Imaging Conference of the Society of Photo-Optical and Instrumentation Engineers 3340–12,128–134 (1998).
7. N. L. Eigler, M. P. Eckstein, K. Mahrer, D. Honig, and J. S. Whiting, “Effect of a stenosis stablilized display on morphological features,” Circ. 96, 1157–1164 (1997).
9. W. B. Pennebaker and J. L. Mitchell, JPEG Still Image Data Compression Standard, (Van Nostrand Reinhold, New York, 1992).
10. A. Zandi, J. Allen, E. L. Schwartz, and M. Boliek, “CREW Lossless/Lossy Medical Image Compression,” IEEE Data Compression Conference, 212–221 (1995).
11. D. M. Green and J. A. Swets, Signal Detection Theory and Psychophysics, Wiley, New York, (1966).
13. M. F. Kijewski, R. G. Swensson, and P. F. Judy, “Analysis of rating data from multiple-alternative tasks,” J. of Math. Psych. 33, 428–451 (1989). [CrossRef]
14. I. Daubechies, Ten Lectures on Wavelets, (SIAM, Philadelphia, 1992). [CrossRef]
15. C. K. Chui, An Introduction to Wavelets, (Academic Press, Boston, (1992).