1352
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 12, DECEMBER 2001
A Multistage Perceptual Quality Assessment for Compressed Digital Angiogram Images Joonmi Oh, Member, IEEE, Sandra I. Woolley*, Member, IEEE, Theodoros N. Arvanitis, Member, IEEE, and John N. Townend
Abstract—This paper describes a multistage perceptual quality assessment (MPQA) model for compressed images. The motivation for the development of a perceptual quality assessment is to measure (in)visible differences between original and processed images. The MPQA produces visible distortion maps and quantitative error measures informed by considerations of the human visual system (HVS). Original and decompressed images are decomposed into different spatial frequency bands and orientations modeling the human cortex. Contrast errors are calculated for each frequency and orientation, and masked as a function of contrast sensitivity and background uncertainty. Spatially masked contrast error measurements are then made across frequency bands and orientations to produce a single perceptual distortion visibility map (PDVM). A perceptual quality rating (PQR) is calculated from the PDVM and transformed into a one to five scale, PQR1 5 , for direct comparison with the mean opinion score, generally used in subjective ratings. The proposed MPQA model is based on existing perceptual quality assessment models, while it is differentiated by the inclusion of contrast masking as a function of background uncertainty. A pilot study of clinical experiments on wavelet-compressed digital angiogram has been performed on a sample set of angiogram images to identify diagnostically acceptable reconstruction. Our results show that the PQR1 5 of diagnostically acceptable lossy image reconstructions have better agreement with cardiologists’ responses than objective error measurement methods, such as peak signal-to-noise ratio A Perceptual thresholding and CSF-based Uniform quantization (PCU) method is also proposed using the vision models presented in this paper. The vision models are implemented in the thresholding and quantization stages of a compression algorithm and shown to produce improved compression ratio performance with less visible distortion than that of the embedded zerotrees wavelet (EZWs). Index Terms—Digital angiography, human visual system, perceptual quality, wavelet compression.
I. INTRODUCTION
C
ORONARY artery disease is the major cause of premature death in the United Kingdom [1]. X-ray cardioangiograms are used to observe coronary blood flow, diagnose arterial disease, and perform coronary angioplasty or bypass surgery. Such images are increasingly represented in digital Manuscript received January 19, 2000; revised September 19, 2001. The Associate Editor responsible for coordinating the review of this paper and recommending its publication was A. Manduca. Asterisk indicates corresponding author. J. Oh and T. N. Arvanitis are with Electronic, Electrical and Computer Engineering, The University of Birmingham, Birmingham B15 2TT, U.K. *S. I. Woolley is with Electronic, Electrical and Computer Engineering, The University of Birmingham, Birmingham B15 2TT, U.K. (e-mail:
[email protected]). J. N. Townend is with the Department of Cardiovascular Medicine, Queen Elizabeth Hospital, The University of Birmingham, Birmingham B15 2TT, U.K. Publisher Item Identifier S 0278-0062(01)11144-4.
form [2], [3]. The digital storage and transmission of cardiovascular images has a significant potential to improve patient care. For example, digitally represented images enable electronic archiving, network transmission and useful manipulation of diagnostic information [4]. A single patient digital angiogram video typically requires 7.5 Mbytes/s for 512 512 8-bit resolutions at 30 frames/s, resulting in 0.25 Mbytes/frame. Data compression is therefore essential for efficient and economic transmission and archiving. In 1994, the Ad Hoc Angiographic Committee adopted a lossless compression method (lossless JPEG) and selected CD-R media for digital storage [5]. Lossless compression methods enable exact reconstruction, but with compression ratios typically less than 2:1. It would take, for example, 6.5 hours to transmit 5 min of 2:1 compressed angiogram video across ISDN6 (364 kbps) lines. Several best case transmission times for uncompressed digital angiogram video across conventional transmission lines are listed in Table I. The effect of lossy JPEG compression on X-ray cardio-angiograms have been investigated and compression ratios of 6–15:1 have been shown to be acceptable for the detection of common diagnostic features, for example, stenosis, calcification, stent, dissection, and bypass grafts [6]–[8]. The application of high-quality lossy compression methods would be acceptable if they could enable perceptually lossless image reconstructions without any loss in diagnostic content for specific tasks, for example, detection and/or recognition of disease. Reliable measures of distortion, introduced by the compression and decompression process, which produce numerical measures of quality are important in the design of compression methods, since the objective of compression methods is to achieve the best compression performance while minimizing distortion [9]. Image quality metrics are important performance variables for digital imaging systems, used to measure the visual quality of compressed images. Quality metrics would, ideally, be able to predict visually lossless compression points as well as provide a perceptually meaningful scale when distortions are significantly above the visual threshold. There are three major types of quality measurements; objective, subjective and perceptual measurement. • Objective quality metrics provide mathematical deviations between original and processed images, such as a mean-squared error (mse)1 or a peak signal-to-noise ratio (PSNR2 ), and are easy to calculate. However, they are
0
2
1mse = [I (i; j ) I (i; j )] =N N ,where I (i; j ) is reference image and I (i; j ) is decompressed image, and N N is the dimensions (row, col) of the reference image. 2PSNR = 10 log (2 1) =mse, where b is the required bit to present a pixel value.
0278–0062/01$10.00 © 2001 IEEE
0
2
OH et al.: MPQA FOR COMPRESSED DIGITAL ANGIOGRAM IMAGES
1353
TABLE I BEST CASE TRANSMISSION TIMES FOR UNCOMPRESSED DIGITAL ANGIOGRAM VIDEO ACROSS CONVENTIONAL TRANSMISSION LINES
poor predictors of distortion visibility since their implementations do not consider human visual sensitivities. Furthermore, they do not accurately predict visual quality across a set of images with large luminance variations or with varying content, such as edges and textured regions [10]. • Subjective quality measures provide numerical values that quantify viewers’ satisfaction with reconstructed images [11]. However, these methods involve time-consuming experiments and observer responses can vary significantly. • Perceptual quality measures are based on models of human visual perception, and can be divided into two categories; image discrimination models and task performance based models. Several perceptual image discrimination quality met— rics have been proposed as alternatives to objective metrics; for example, methods which incorporate luminance adaptation and contrast sensitivity functions (CSF) [12], metrics which incorporate observer performances for supra-threshold artifacts [13], and threshold perceptual metrics [14]. The perceptual metrics can provide more consistent estimation of image quality than objective metrics when artifacts are near the visual threshold. Image discrimination models used in perceptual quality assessment, however, have been developed for measuring general quality degradation introduced by compression processes. The implementation of these metrics is also often complex, and time-consuming subjective psychophysical testing is required for validation [15]. Task-based model observers have been designed to — predict human visual detectability of a signals embedded in noisy backgrounds [16]. Recent studies involving the detection of simulated lesions in single frame JPEG compressed X-ray cardio-angiograms shows that model observers can predict the effect of compression on human performance [17], [18]. The effect of quality degradations on the performance of detecting diagnostic features for medical image requires further investigation. Receiver operating characteristic (ROC) analysis has been widely used for measuring diagnostic accuracy for computerprocessed medical images employing. The ROC analysis is based on the observers’ rating of each image representing their subjective confidence level that the image does or does not contain evidence of disease (or abnormality) [19]. The ROC curve represents the tradeoff between the true and false positive fraction as the observer varies the decision threshold. ROC methods are appropriate for detection tasks, however,
there are some fundamental problems in their application to images which require diagnosis via the detection of localized and/or multiple abnormalities [20]. In addition, many common diagnostic tasks involve more than simple binary “yes/no” (abnormal/normal) decisions. In this paper, we propose a perceptual thresholding and CSFbased uniform quantization (PCU) method and a multistage perceptual quality assessment (MPQA) model for test set of angiogram images. The PCU is human visual system (HVS)-informed perceptual thresholding and quantization method. The MPQA model is an objective measure of subjective quality which employs models of the HVS for the measurement of image distortions, introduced by the compression process. The MPQA could be used to optimize the compression rate/distortion trade-off. The organization of the paper is as follows. Section II describes the implementation of different thresholding and quantization strategies, including the proposed PCU and an embedded zerotree wavelets (EZWs) implementation, on a test set of angiogram images to introduce information losses. Section III contains the detail description of the proposed MPQA model. Section IV describes the compression results, and compares the performance of objective metrics and the proposed MPQA model, as well as clinical quality testing. Finally, Section V provides conclusions. II. THE COMPRESSION PROCESS Wavelet transformations compact image energy into smaller numbers of coefficients and have good localization characteristics in the spatial-frequency domain. Such characteristics are ideal for compression [21]. Reversible biorthogonal wavelet filters are considered in this paper and described more in detail in [21]–[23]. The Perceptual thresholding and CSF-based uniform quantization (PCU) and EZWs were evaluated for lossy wavelet implementation to investigate the perceptual effects of information loss. We focus on the identification of acceptable bit-rates (or lossiness) of compressed angiogram images, from a range of lossy reconstruction, to ensure sufficient accuracy for diagnostic purposes (i.e., detecting coronary artery lesions). A simple coefficient thresholding process can be performed by setting specific wavelet transformed detail coefficients to zero while retaining most of the energy of the original image. While easy to implement, this method fails to consider visual sensitivities at different resolutions and orientations in the wavelet domain. Hence, resolution- and orientation-dependent perceptual thresholding was implemented. Fourier transformed first-level wavelet detail subbands were geometrically rearranged as shown in Fig. 1 and then sum of cortex filters [24] of six orientations (resulting in a donut ring shape as shown in
1354
Fig. 1.
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 12, DECEMBER 2001
A block diagram of the perceptual thresholding and CSF-based uniform quantization (PCU) processes.
where is the display factor, is contrast sensitivity from the is gain from the quantization domain. The display CSF, and factor, , is display modulation transfer function at the center frequency of band [25], and set to one as an ideal case. The gain from the quantization domain of the reconstructed spatial , was set to one for the lowest level of wavelet deimage, 3) and increased as a factor of composed frequency band ( 2 ) since a two for the higher frequency band (i.e., factor of two gain is observed from previous study [26]. The CSF can be used to maximize quantization errors in different bands, yet ensure these errors are not detectable, thus achieving visually lossless compression [25]. Scanned coefficients were then entropy encoded with run-length and Huffman encoding. Finer quantization could be applied to preserve information in the regions of interest (ROIs). For comparison, the EZW was also evaluated [27], [28]. III. MPQA
Fig. 2. CSF model response and central frequency sensitivity sampling method for wavelet subbands.
Fig. 1) were applied. As a result, diagonal orientation details were thresholded more than their horizontal and vertical counterparts. The Daly CSF model describes sensitivity variations as a function of radial spatial frequency, light adaptation level, image size, and lens accommodation due to distance, orientation and eccentricity [14]. Daly’s model has a bandpass characteristic and anisotropic behavior caused by lack of sensitivity in the region of 45 . The CSF model response and a central frequency sensitivity sampling method for wavelet subbands are shown in Fig. 2. Sampled contrast sensitivity, , is used to define undetectable contrast errors at the quantization stage. Frequency band and orientation-dependent sensitivity from the CSF and display factors were considered to determine an appropriate quantization step size for frequency band ( 1, 2, and 3 where 1 and 3 represents high- and low-frequency bands, respectively), , and defined as [25] (1)
Many existing image quality models are based on distance metrics between original and processed images. For example, Daly’s visual differences predictor (VDP) is designed to determine the degree to which physical differences (i.e., incorrect luminance and chrominance) become visible, rather than mathematical differences (i.e., incorrect code values) [29]. Several perceptual image discrimination quality assessment models are compared in [30]. As shown in Fig. 3, the implemented MPQA method models the HVS as a series of stages, including amplitude nonlinearity, octave bandwidth spatial frequency decompositions into six orientations using Watson’s cortex transformation [24], and contrast masking based on CSF modeling and region classification from the decomposed images. A perceptual distortion visibility map (PDVM) is then produced via a distance computation and summation of errors across different spatial frequency bands. A single quantitative perceptual quality rating (PQR) is then calculated from the distortion map converting fidelity to quality. The proposed model is similar to other perceptual quality assessment models (for example, Daly’s VDP model [14]), while it is differentiated by the inclusion of contrast masking as a function of background uncertainty. The details of which are presented in Section III-B.
OH et al.: MPQA FOR COMPRESSED DIGITAL ANGIOGRAM IMAGES
1355
Fig. 3. A block diagram of the MPQA model.
Fig. 4. The human eye’s normalized amplitude nonlinearity response.
A. Amplitude Nonlinearity Since human eyes are less sensitive to noise in brighter luminance backgrounds, amplitude nonlinearity is an important consideration. The human eye’s normalized amplitude nonlinearity response is shown in Fig. 4 and modeled as [14] (2) where output,
is the normalized amplitude nonlinearity model is the input luminance value .
B. Spatial Frequency Decomposition and Contrast Masking The cortex transformation [24], introduced by Watson, is based on a combination of psychophysically measured spatial frequency selectivity and the measurements of two-dimensional
cortex receptive fields. Cortex transformations can be used to model the visual cortex located at the back of the brain which tunes to specific frequencies and orientations. Ideally, perceptual image quality metrics should be able to characterize spatial variations in quality across an image, since the visibility of artifacts is highly dependent on their location with an image. Advantages of the cortex transformation include its flexibility and invertibility. Flexibility comes from its modular construction, enabling variation of parameters such as bandwidths and specific shapes of the radial frequency and orientation responses. Nonorthogonality and computational complexity are disadvantages of the cortex transformation. Cortex filter responses for six frequency bands at 30 orientations, and one – ) of decomposed frequency bands ( for a test angiogram image are shown in Fig. 5(a) and (b), respectively. The wavelet transformation, which is a multiresolution decomposition with space and frequency localization properties, is orthogonal and has self-similarity across resolutions. More importantly, the independent octave-wide oriented frequency bands can allow wavelet-based methods to take advantage of the masking characteristics of the visual system. In wavelet decompositions, however, diagonal bands for ) are contained within orientations near 45 and 135 (or the same band which does not occur in the visual system. After spatial frequency decomposition, the global contrast of each , is calculated as frequency band,
and
(3)
is the value of -frequency band and -orientawhere is the mean value of baseband (the tion cortex filtered image, 6) image. lowest frequency band where Masking quantifies the effect of decreased visibility as a function of contrast and uncertainty created by background [31]. Both CSF and the spatial frequency selectivity of the visual system play key roles in modeling the specifics of masking. Contrast masking is modeled as a threshold elevation process
1356
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 12, DECEMBER 2001
Fig. 5. (a) Cortex filter responses, CX(f ; ), for six frequency bands at 30 orientations. f (k = 1,2, . . ., 6) is the frequency band, where k = 1 and k = 6 represents high- and low-frequency residue, respectively. is the angle in degree measured from the horizontal axis. (b) A decomposed frequency band of a test angiogram image in six orientations using cortex filters in (a).
as shown in Fig. 6. It is shown that the low uncertainty stimuli exhibit a threshold elevation curve with 0.7 log-log slopes, while high uncertainty stimuli would have 1 log-log slopes [32], [33]. This implies that the human eye can tolerate larger errors in textured areas than along edges of the same contrast. Masking therefore adjusts the threshold of contrast error detection depending on the local image content. The spatially decomposed images are therefore classified into flat, edge and texture regions to consider relationship between stimulus and background uncertainty. Model performance improvement introduced by spa-
tial frequency masking has been shown in [34] and [35]. An original test angiogram image and its classified image are shown in Fig. 7(a) and (b), respectively. Areas having lower contrast than the base threshold contrast given by CSF are classified as flat regions. A center-frequency sampling strategy from CSF, giving an average sensitivity for each frequency band, is used to calculate the base threshold for -frequency band -orientation, , by computing the sensitivity inverse. The base sensitivity values of one quarter of the total frequency band are sampled from the CSF and values for the remaining three quarter
OH et al.: MPQA FOR COMPRESSED DIGITAL ANGIOGRAM IMAGES
1357
C. The Perceptual Visibility Distortion Map and Perceptual Quality Rating Minkowski error summation [36] is employed across different frequency bands and orientations to produce a single PDVM, , as (6) 4 [37]. Minkowski summation over all pixels where into a single quantitais calculated by converting tive perceptual quality rating, PQR, and converting fidelity to quality as (7) Fig. 6. Contrast masking performed by threshold elevations.
3 [38]. Finally, the PQR value is transformed into where a scale from one to five for direct comparison with the mean opinion score scale, typically used for subjective ratings, defined as (8) where
0.8 [35]. IV. RESULTS
Fig. 7. (a) An original angiogram image (Angiogram 1) and (b) Classified flat, edge and texture regions for (a).
frequency bands are achieved considering symmetry properties of the CSF and cortex filters. Edge areas are detected by using a Sobel edge detector. Remaining regions are classified as texture regions. Masking is implemented by threshold elevation as is [35], see (4) at the bottom of the page, where the detection threshold in the presence of a masker for -freis the base threshold for quency band -orientation, -frequency band -orientation, is the contrast of 0.7 for edge area, 1 for texture area. masker, The threshold value for each frequency band and location, , is then used to determine whether the contrast error between the original and processed image is visible. The contrast error visibility of each frequency band and orientation, , is defined as (5) is the original image contrast and is where the processed image contrast for -frequency band -orientation.
Nine diagnostically important frames/regions of interest (F/ROIs) were selected from seven patients’ digital angiogram video sequences by an expert cardiologist. Different thresholding and quantization strategies (as described in Section II) were implemented on this test set. Four different image reconstructions were achieved for each of the nine test angiogram images, resulting in a range of compression performances; 1.47 0.20 bpp for the PCU method, and 4.37 0.22 bpp for the EZW method. Two lower bit-rates of image reconstruction between two compression methods matched well while two higher bit-rates resulted in significantly higher3 for the EZW than PCU method, because the perceptual thresholding in the PCU worked efficiently and it was difficult to determine performance to match with the EZW output. However, the visual quality of the higher bit-rates reconstruction between two compression methods was similar and calculated scores of them matched well. PQR A. Compression Method Performances Resolution- and orientation-dependent thresholding were implemented by using a sum of cortex filters as shown in Fig. 1. A test angiogram image (Angiogram 2), its dc coefficient centered 3Double tailed unpaired T -tests were performed as the test of significance unless otherwise specified.
(otherwise)
(4)
1358
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 12, DECEMBER 2001
Fig. 9. The progressive reconstruction of an angiogram image using EZW.
Fig. 8. (a) An original angiogram image (Angiogram 2), (b) Fourier transformed first level wavelet detail subbands, (c) Geometrical rearrangement of subbands, and (d) Perceptually thresholded subbands. TABLE II PERCENTAGE OF RETAINED ENERGY AND COEFFICIENTS SET TO ZERO (PZ) FOR ALL LEVEL WAVELET SUBBANDS AND THE 1ST LEVEL SUBBANDS OF A TEST ANGIOGRAM IMAGE (ANGIOGRAM 1) AFTER PERCEPTUAL THRESHOLDING Fig. 10. Bit-rates versus PSNR for the test angiogram image set (PCU method).
and refinement. Fig. 9 shows the progressive reconstruction of a test angiogram image (Angiogram 1) using the EZW algorithm. B. Objective Quality Assessment Versus MPQA
Fourier transformed first level wavelet subbands, their geometrical rearrangement, and perceptually thresholded subbands are shown in Fig. 8(a)–(d), respectively. The percentage of retained energy (calculated by using the second-order vector norms) and coefficients set to zero for all level wavelet subbands and the first level subbands of a test angiogram image (Angiogram 1) after perceptual thresholding are listed in Table II. As shown in Table II, diagonal orientation details are thresholded more than their horizontal and vertical counterparts, due to the decreased sensitivity of the HVS for diagonal orientation. By increasing the quantization step size determined as a function of the HVS contrast sensitivity, coarser quantization was achieved resulting in lower bit-rates. Zerotree coding is based on the hypothesis that if a wavelet coefficient at a coarse scale is insignificant with respect to a given threshold, , then all coefficients at the same orientation in the same spatial location at finer scales are very likely to be also insignificant with respect to . A zerotree root is encoded with a special symbol indicating that the whole tree is insignificant. The EZW coding is performed by significance map coding
The analysis of error images (differences between original and decompressed images) provides coarse approximations of the quality perceived by viewers, providing mathematical deviations between original and decompressed images. It was found that the EZW produced statistically significant higher PSNR , particuvalues than the PCU at the similar bit-rates larly evident at lower bit-rates. For compression/decompression systems objective quality metrics, such as PSNR and mse, can be used to perform the rate-distortion trade-off between image quality and compression ratio performance. Objective quality metrics, however, are not always reliable predictors for expected bit-rates. For example, compressed bit-rates fall across a wide range of PSNR values, and vice versa, for the PCU method as shown in Fig. 10. One of the motivations for the development of a perceptual quality measurement is to measure (in)visible physical differences between original and processed images as described in Section III. The implementation block diagram for the perceptual quality measurement was shown in Fig. 2. The PDVMs of two differently compressed images for an Angiogram 1 are shown in Fig. 11. As shown in Fig. 11, the EZW produced larger perceptual distortions than the PCU
OH et al.: MPQA FOR COMPRESSED DIGITAL ANGIOGRAM IMAGES
1359
Fig. 11. PDVMs for a test angiogram image (angiogram 1) (lighter areas indicate larger perceptual distortion). (a) PCU method. (b) EZW method.
Fig. 12.
The graphical user interface configuration used for diagnostic “quality boundary” clinical testing.
method at similar bit-rates. The PDVM was converted into a single number, referred to as the PQR , scaled from one to five, representing Bad, Poor, Fair, Good and Excellent, respecscores resulted in a range of 2.9–4.6 tively. Calculated PQR (1.47 0.20 bpp) for the PCU method, and 2.2–4.4 (4.37 0.22
bpp) for the EZW method for test angiorgram data set. It was observed that the PCU method produced higher PQR , particularly at lower bit-rates, than the EZW at similar bit-rates since the PCU incorporated properties of the HVS, however, it was 0.294). not statistically significant (
1360
Fig. 13.
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 12, DECEMBER 2001
Diagnostically acceptable maximum, average and minimum bit-rates for test angiogram images for three cardiologists. (a) PCU method. (b) EZW method.
C. Clinical Quality Testing of Compressed Angiogram Images The decompressed test images were then used in subjective diagnostic quality assessments with a simple Graphical User Interface (GUI) as shown in Fig. 12. Three cardiologists (A, B, and C) were asked to examine reconstructed image quality and consider diagnostic responsibility when selecting the “lossiest” acceptable image reconstructions. The decompressed test angiogram image set were displayed with increased/decreased information lossiness at a rate of 0.6 frames/s, and experts were asked to identify “quality boundaries” of diagnostically acceptable/unacceptable image representations. Fig. 12 shows the GUI configuration developed for this test. The GUI enabled observers to view frames in a movie format at 0.6 frame/s or frame-by-frame with forward (“ ”) and backward (“ ”) control functions. This functionality was implemented at the request of observing cardiologists. It was observed that cardiologists first identified ROIs and focused on the quality degradation within these regions. Images were displayed on a 14 in LCD and the test was performed in room light. Viewers were allowed to repeat the display of each image sequence if they were unable to make a selection after the first pass. Four different qualities of image reconstructions were presented for each of the nine test angiogram images arranged in order of the PCU and EZW. The results of the diagnostically acceptable “quality boundary” test for maximum, average and minimum bit-rates of the nine test angiogram images for two compression methods for the three cardiologists are shown in Fig. 13. The bit-rate, of the lossiest diagnostically acceptable PSNR and PQR frames for two compression methods selected by three cardiologists are listed in Table III. The selected bit-rates, PSNR among three cardiologists were not significantly and PQR different, except between cardiologist B and C for the PCU 0.001 for the bit-rates and 0.009 for the method ( PQR ) (See Table III). It was found that the lossiest diagnostically acceptable frames had significantly lower average for the PCU (0.87 bpp) than for the EZW bit-rates (2.03 bpp), indicating an improved performance for the proposed compression method. It was also found that the PQR
TABLE III THE BIT-RATE, PSNR AND PQR OF SELECTED THE LOSSIEST DIAGNOSTICALLY ACCEPTABLE FRAMS FOR TWO COMPRESSION METHODS BY THREE CARDIOLOGISTS ( INDICATES SIGNIFICANT DIFFERENCES (P < 0:01) BETWEEN TWO OBSERVERS’ SELECTION)
scores of selected frames for two compression methods were 0.156) among three cardiologists, while the PSNR similar ( , showing the values were significantly different superiority of the proposed MPQA model over the PSNR, for the prediction of diagnostically acceptable quality degradation. The trial test sample of angiogram images used in our study was necessarily limited, however, a more substantial set would be desirable for further investigation. V. DISCUSSION AND CONCLUSION Diagnostically important digital angiogram images were compressed using different thresholding and quantization strategies; the PCU and EZW, to investigate the perceptual effects
OH et al.: MPQA FOR COMPRESSED DIGITAL ANGIOGRAM IMAGES
of information loss. Different methods resulted in a range of bit-rates. The proposed MPQA model, which generates visible distortion maps and quantitative error measures informed by considerations of the HVS, was implemented. We focused on the identification of cardiologists’ assessment of diagnostically acceptable bit-rates (or lossiness) of compressed angiogram images to ensure sufficient accuracy. It was found that the lossiest acceptable reconstruction of the proposed PCU method resulted in significantly lower average bit-rates than the EZW. It was also found that the implemented perceptual quality rating (PQR ) of diagnostically acceptable lossy image reconstructions has better agreement with cardiologist responses for two compression methods than the objective error measurement methods (i.e., PSNR). It indicates that the usefulness of the proposed model to assess image quality reconstructed by different compression methods, particularly having large variation of objective quality metric or bit-rates. However, the relationships between cardiologists’ rating and the MPQA model, and between the MPQA and task-based performance remain open questions for further work. The effect of quality degradations on the performance of detecting diagnostic features for medical image also requires further investigation. Task-based measures, for example, ROC studies involving physician-detection of lesions, could usefully address whether diagnostic performance is also improved in the PCU method over the EZW method. The human vision and quality assessment models presented in this paper may be useful in the design of compression algorithms for economical storage, fast transmission of images and optimal compression rate/distortion trade-off. REFERENCES [1] E. D. Grech and D. R. Ramsdale, Practical Interventional Cardiology, London, U.K.: Martin Dunitz Ltd., 1997. [2] B. F. A. Van Meurs, “Information management in the cardiology department—an analysis of current options for replacing cinefilm,” Int. J. Cardiac Imag., vol. 11, Suppl 3, pp. 159–163, 1995. [3] J. T. Cusma, L. A. Spero, J. D. Hanemann, T. M. Bashore, and K. G. Morris, “A multiuser environment for the display and processing of digital cardiac angiographic images,” Proc. SPIE, vol. 1233, pp. 310–320, 1990. [4] H. K. Huang, PACS: Picture Archiving and Communication Systems in Biomedical Imaging. New York: VCH, 1996. [5] P. B. Condit, “Requirements for cardiac interchange media and the role of recordable CD,” Int. J. Cardiac Imag., vol. 11, Suppl 3, pp. 153–157, 1995. [6] R. A. Kerensky et al., “American college of cardiology/European society of cardiology international study of angiographic data compression phase I,” Eur. Heart J., vol. 2000, pp. 668–678, 2000. [7] M. P. Eckstein, J. L. Bartroff, C. A. Morioka, D. J. Vodopich, and J. S. Whiting, “Feature stabilized digital X-ray coronary angiogams improve human visual detection in JPEG compressed images,” Optics Express, vol. 4, no. 6, pp. 193–199, Mar. 1999. [8] V. H. Rigolin et al., “Compression of digital coronary angiograms does not affect visual or quantitative assessment of coronary artery stenosis severity,” Amer. J. Cardiol., vol. 78, pp. 131–135, 1996. [9] P. C. Cosman, R. M. Gray, and R. A. Olshen, “Evaluating quality of compressed medical images: SNR, subjective rating, and diagnostic accuracy,” Proc. IEEE, vol. 82, pp. 919–932, June 1994. [10] B. Girod, “What’s wrong with mean-squared error?,” in Digital Images and Human Vision, A. B. Watson, Ed. Cambridge, MA: MIT Press, 1993, pp. 207–220. [11] Methodology for the Subjective Assessment of the Quality of Television Pictures, 1994. ITU-R, ITU-R Recommendation 500-6. [12] A. J. Ahumada Jr., “A simple vision model for inhomogeneous image quality assessment,” presented at the SID Digest, vol. 29, 1998, p. Paper 40.1.
1361
[13] M. Miyahara, K. Kotani, and V. R. Algazi, “Objective picture quality scale (PQS) for image coding,” IEEE Trans. Commun., vol. 46, pp. 1213–1226, Sept. 1998. [14] S. Daly, “The visual differences predictor: an algorithm for the assessment of image fidelity,” in Digital Images and Human Vision, A. B. Watson, Ed. Cambridge, MA: MIT Press, 1993, pp. 179–206. [15] M. P. Eckert and A. P. Bradley, “Perceptual quality metrics applied to still image compression,” Signal Processing, vol. 70, pp. 177–200, 1998. [16] H. H. Barrett, J. Yao, J. P. Rolland, and K. J. Myers, “Model Observers for Assessment of Image Quality,” in Proc. Nat. Acad. Sci. USA, vol. 90, 1993, pp. 9758–9765. [17] M. P. Eckstein, C. A. Abbey, F. O. Bochud, J. L. Bartroff, and J. S. Whiting, “The effect of image compression in model and human performance,” Proc. SPIE, vol. 3663, pp. 243–252, 1999. [18] C. A. Morioka et al., “Observer performance for JPEG versus wavelet image compression of X-ray coronary angiograms,” Optical Express, vol. 5, no. 1, pp. 8–19, July 1999. [19] J. A. Swets, Signal Detection Theory and ROC Analysis—Psychology and Diagnostics: Collected Papers. Mahwah, NJ: Lawrence Erlbaum Associates, 1996. [20] D. P. Chakraborty and L. H. L. Winter, “Free-response methodology: Alternate analysis and a new observer-performance experiment,” Radiology, vol. 174, no. 3, pp. 873–881, Mar. 1990. [21] G. Strang and T. Nguyen, Wavelets and Filter Banks. Cambridge, MA: Wellesley-Cambridge Press, 1997. [22] M. J. Gormish, E. L. Schwartz, A. Keith, M. Boliek, and A. Zandi, “Lossless and nearly lossless compression for high-quality images,” Proc. SPIE, vol. 3025, pp. 62–70, 1997. [23] J. Oh, S. I. Woolley, T. N. Arvanitis, and J. N. Townend, “Investigation of quality assessments for wavelet-compressed digital angiogram images,” Proc. SPIE, vol. 3658, pp. 436–447, 1999. [24] A. B. Watson, “The cortex transform: rapid computation of simulated neural images,” Comput. Vis., Graphic., Image Processing, vol. 39, no. 3, pp. 311–327, 1987. [25] P. W. Jones, S. Daly, R. S. Gaborski, and M. Rabbani, “Comparative study of wavelet and DCT decompositions with equivalent quantization and encoding strategies for medical images,” Proc. SPIE, vol. 2431, pp. 571–582, 1995. [26] L. Vandendorpe, “Optimum quantization for image subband coders,” Signal Processing:Image Commu., vol. 4, no. 1, pp. 65–79, Nov. 1991. [27] J. M. Shapiro, “Embedded image coding using zerotrees of wavelet coefficients,” IEEE Trans. Signal Processing, vol. 41, pp. 3445–3462, Dec. 1993. [28] A. Munteanu, J. Cornelis, and P. Cristea, “Wavelet-based lossless compression of coronary angiographic images,” IEEE Trans. Med. Imag., vol. 18, pp. 272–281, Mar. 1999. [29] S. J. Daly, “Method and apparatus for determining visually perceptible differences between images,” U.S. Patent, 5 394 483, 1995. [30] R. Eriksson, B. Andren, and K. Brunnstrom, “Modeling the perception of digital images: A performance study,” Proc. SPIE, vol. 3299, pp. 88–97, 1998. [31] W. Osberger, N. Bergmann, and A. J. Maeder, “An automatic image quality assessment technique incorporating high-level perceptual factors,” ICIP, vol. 1, pp. 4–7, 1998. [32] J. M. Foley and G. M. Boynton, “A new model of human luminance pattern vision mechanism: analysis of the effects of pattern orientation, spatial phase and temporal frequency,” Proc. SPIE, vol. 2054, pp. 32–42, 1994. [33] M. A. Losada and K. T. Mullen, “Color and luminance spatial tuning estimated by noise masking in the absence of off-frequency looking,” J. Opt. Soc. Amer. A, vol. 12, no. 2, pp. 250–260, 1995. [34] M. P. Eckstein, A. Ahumada, and A. B. Watson, “Image discrimination models predict signal detection in natural medical image backgrounds,” Proc. SPIE, vol. 3016, pp. 44–56, 1997. [35] W. Osberger, “Perceptual Vision Models for Picture Quality Assessment and Compression Applications,” Ph.D. dissertation, Sch. Elect. Electron. Syst. Eng., Queensland Univ. Technol., Queensland, Australia, 1999. [36] L. A. Olzak and J. P. Thomas, “Seeing spatial patterns,” in Handbook of Perception and Human Performance, K. R. Boff, L. Kaufman, and J. P. Thomas, Eds. New York: Wiley, 1986. [37] A. B. Watson, “Temporal sensitivity,” in Handbook of Perception and Human Performance, K. R. Boff, L. Kaufman, and J. P. Thomas, Eds. New York: Wiley, 1986. [38] R. Hamberg and H. de Ridder, Time-Varying Image quality: Modeling the relation between instantaneous and overall quality, vol. 1234, 1998. IPO manuscript.
