You are on page 1of 6

Eur Radiol (2007) 17: 15291534 DOI 10.

1007/s00330-006-0509-6

COMPUTER APPLICATI ONS

Kyoung Ho Lee Young Hoon Kim Bo Hyoung Kim Kil Joong Kim Tae Jung Kim Hyuk Jung Kim Seokyung Hahn

Irreversible JPEG 2000 compression of abdominal CT for primary interpretation: assessment of visually lossless threshold

Received: 23 May 2006 Revised: 16 September 2006 Accepted: 12 October 2006 Published online: 22 November 2006 # Springer-Verlag 2006

H. J. Kim Department of Radiology, Seoul Medical Center, Samsung-dong, Gangnam-gu, Seoul, 171-1, Korea

This study was supported by Seoul R&BD Program, Republic of Korea (project number, 10675). K. H. Lee . Y. H. Kim (*) . B. H. Kim . K. J. Kim . T. J. Kim Department of Radiology, Seoul National University Bundang Hospital, 300 Gumi-dong, Bundang-gu, Seongnam-si, Gyeonggi-do, 463-707, South Korea e-mail: yhk@snubhrad.snu.ac.kr Tel.: +82-31-7877609 Fax: +82-31-7874011 S. Hahn Medical Research Collaborating Center, Seoul National University Hospital, 28 Yongon-dong, Chongno-gu, Seoul, 110-744, South Korea

Abstract To estimate the visually lossless threshold for Joint Photographic Experts Group (JPEG) 2000 compression of contrast-enhanced abdominal computed tomography (CT) images, 100 images were compressed to four different levels: a reversible (as negative control) and irreversible 5:1, 10:1, and 15:1. By alternately displaying the original and the compressed image on the same monitor, six radiologists independently determined if the compressed image was distinguishable from the original image. For each reader, we compared the proportion of the compressed images being rated distinguishable from the original images between the reversible compression and each of the three irreversible compressions using the exact test for paired propor-

tions. For each reader, the proportion was not significantly different between the reversible (01%, 0/100 to 1/100) and irreversible 5:1 compression (03%). However, the proportion significantly increased with the irreversible 10:1 (9599%) and 15:1 compressions (100%) versus reversible compression in all readers (P<0.001); 100 and 95% of the 5:1 compressed images were rated indistinguishable from the original images by at least five of the six readers and all readers, respectively. Irreversibly 5:1 compressed abdominal CT images are visually lossless and, therefore, potentially acceptable for primary interpretation. Keywords Data compression . Computed tomography . Teleradiology . Computer storage devices

Introduction
Irreversible (lossy) image compression appears to be an immediate and effective means to cope with the computed tomography (CT) data explosion [14]; however, such compression techniques are not always accepted by radiologists. One of the reasons for this reluctance is the difficulty in establishing an acceptable compression level that does not result in the deterioration of image quality to the point where necessary diagnostic information is lost.

The acceptable threshold is dependent on several independent parameters, including image content, compression algorithm, and specific reading task [5, 6]. For instance, it has been reported that detection performance of CT for liver nodules is preserved with up to 10:1 compression [7]; however, it is uncertain from this study if this compression level is acceptable for the characterization of the nodules and for the detection of a potential auxiliary or coincidental finding which might be clinically important in the same CT dataset.

1530

If a compressed image cannot be distinguished from the original (noncompressed) image by radiologists, there is no basis for arguing that this visually lossless compression hinders any diagnostic accuracy [8, 9]. In other words, a visually lossless threshold for image compression can exist at a certain point between a mathematically lossless threshold (reversible threshold) and a diagnostically lossless threshold. The purpose of this study was to estimate the visually lossless threshold for Joint Photographic Experts Group (JPEG) 2000 compression of abdominal CT images. We have chosen the visually lossless threshold rather than the diagnostically lossless threshold for a specific reading task because we believe the former would be more conservative and widely accepted, potentially even for primary interpretation.

By generating a random sequence for the reconstructed images in each patient, one image per patient was selected to form a 100-image set. These images included 48 sections at or above the umbilicus and 52 sections below the umbilicus. Image compression Each of the 100 original images had a bit depth of 12 bits/ pixel packed into two bytes. Using a two-dimensional JPEG 2000 wavelet-based image compression algorithm (Pegasus Imaging Co., Tampa, FL, USA), each original image was compressed to four different levels: a reversible (as negative control) and irreversible 5:1, 10:1, and 15:1. These compressed images were then decompressed, yielding 400 compressed (and then decompressed) images for comparison with the corresponding original images. We used the following set of compression parameters: single tile; 6 levels of wavelet decomposition; size of code block 6464; size of precinct 32,76832,768; and a single layer. The actual compression levels (the ratio of original size at 16 bits/pixel to compressed size in bits/pixel) achieved for the four nominal levels were 3.460.28 (meanSD), 5.06 0.27, 10.010.07, and 15.000.16, respectively. The variations from the nominal levels were considered unimportant in this study. Visual analysis Six board-certified body radiologists (including K.H.L. and Y.H.K.) participated in the analysis. They had 611 years (6, 6, 7, 8, 9, and 11 for readers 16, respectively) of working experiences in interpreting abdominal CT findings. Each reader was informed of the purpose of the evaluation, a description of the study protocol, and the structure of analyzed datasets (100 images compressed to the four different levels). Each of the 400 compressed (and then decompressed) images was paired with the corresponding original images for visual comparison. The 400 pairs of the original and compressed images were randomly assigned to eight reading sessions, whilst avoiding the repetition of any patient at a session. The order of reading sessions was changed among readers. Reading sessions were separated by a minimum of 2 weeks. To compare an image pair, we used a method which is similar to that used by Slone et al. [9]. On a single monitor, the reader selectively toggled between the two images in a rapid fashion by using the wheel on the mouse. The original image, which was identified as such, was always displayed first, and the compressed image [the reversibly compressed image (as negative control) or irreversibly compressed image] was displayed later. The reader could return to the first image as desired. Each reader independently determined

Materials and methods


Our Institutional Review Board required neither its approval nor informed patient consent for the use of clinical images in this study, as the patients confidentiality was guaranteed by using anonymized images. CT scanning This study included 100 consecutive adult patients (17 85 years old, 57 males and 43 females) who underwent single-phase contrast-enhanced abdominal CT in Seoul National University Bundang Hospital during a period of 12 days in June 2005. We did not confirm the reasons for CT examination in these patients. Contrast-enhanced abdominal CT examinations were performed using 16-detector-row CT scanners (Brilliance, Philips Medical Systems, Cleveland, OH, USA). Intravenous nonionic contrast material (2 ml/kg, Ultravist 370, Schering, Berlin, Germany) was administered at a rate of 3 ml/s. Bolus-tracking software was used to trigger scanning 60 s after the aortic enhancement reached a 150-HU threshold. Raw projection data were obtained using the following scanning parameters: scan range, from diaphragm to symphysis pubis; detector collimation, 1.5 mm; gantry rotation time, 0.5 s; tube potential, 120 kVp; and pitch, 1.171.25. Tube current was automatically modulated according to the patients body size and the asymmetric nature of the scanned object to give the same level of image noise as a reference image (Dose-Right, Philips Medical Systems, Cleveland, OH, USA). Effective mAs [10] ranged between 121 and 172 (meanSD, 14813). The raw projection data were reconstructed into 5-mm transverse sections at 4-mm intervals. A soft tissue reconstruction algorithm (filter type C) and a matrix of 512512 pixels were used. The field of view ranged between 250 and 369 mm.

1531

if the second image was identical to the first image or if any detectable difference was present (binary response). When making comparisons, the readers were asked to pay attention to structural detail, particularly small vessels and the edge of organs, and the texture of uniform attenuation areas such as the solid organs and soft tissues. Images were displayed in a one-by-one format using a Digital Imaging and Communications in Medicine image viewing software (Pi-view Star version 5.0.7, Infinitt, Phillipsburg, NJ, USA), a flat-panel monochrome monitor (ME315, Totoku, Tokyo, Japan) with a matrix size of 1,5362,048 and a diagonal display size of 20.8 inches (52.8 cm), and a matching video hardware (LV32P1, Totoku, Tokyo, Japan). All annotations and labels suggesting the compression level were toggled off. Images were initially presented with specific window settings (window level, 20 HU; window width, 400 HU), but the readers were encouraged to adjust window centers and level settings. There was no restraint on the reading distance, and the readers were encouraged to magnify images and to inspect the images as closely as needed to discern visual differences. The ambient room light was subdued. Reviewing was conducted at the readers convenience, without a time constraint. Statistical analysis For each reader and for each compression level, a biostatistician (S.H.) calculated the proportion that the compressed images being rated distinguishable from the original images and the corresponding 95% confidence interval [11]. For each reader, the proportions in the irreversible 5:1, 10:1, and 15:1 compressions were compared with that in the reversible compression (as negative control) by using the exact test for paired proportions [12]. A p value of less than 0.05 was considered to indicate a statistically significant difference.

Results
Each reader rated 01% (0/100 to 1/100) of the reversibly compressed images (as negative control) and 03% (0/100 to 3/100) of the irreversibly 5:1 compressed images distinguishable from the corresponding original images; 9599% (95/100 to 99/100) of the 10:1 compressed images and 100% of the 15:1 compressed images were distinguishable from the original images (Table 1, Fig. 1). Four of the six readers (readers 2, 3, 4, and 6) rated none of the reversibly compressed images and irreversibly 5:1 compressed images distinguishable from the original images. Consequently, p values could not be calculated for the comparisons between the reversible and irreversible 5:1 compressions for these readers. For readers 1 and 5, the proportion that the compressed images were rated distinguishable from the corresponding original images was not significantly different between the reversible and irreversible 5:1 compressions (p=0.50 and 0.63, respectively). However, the proportion significantly increased with the irreversible 10:1 and 15:1 compressions versus the reversible compression in all readers (p<0.001). In 86 (86%) of 100 patients, all readers rated the 5:1 compressed image indistinguishable from the original image, and the 10:1 compressed image distinguishable. In nine (9%) patients, all readers rated the 5:1 compressed image indistinguishable; and none to five readers rated the 10:1 compressed image distinguishable. In five (5%) patients, five readers rated the 5:1 compressed image indistinguishable, and all readers rated the 10:1 compressed image distinguishable. Therefore, 100 (100/100) and 95% (95/100) of the 5:1 compressed images were rated indistinguishable from the original images by at least five of the six readers and all readers, respectively.

Table 1 Results of visual analysis of 100 abdominal CT imagesa Reader Compression level Reversible Irreversible 5:1 1 2 3 4 5 6
a

10:1 (0.3, 7.7) (0, 4.6) (0, 4.6) (0, 4.6) (0.8, 9.2) (0, 4.6) 98 99 96 97 95 98 (92.3, (93.8, (89.5, (90.9, (88.2, (92.3, 99.7) 100.0) 98.7) 99.2) 98.1) 99.7)

15:1 100 100 100 100 100 100 (95.4, (95.4, (95.4, (95.4, (95.4, (95.4, 100) 100) 100) 100) 100) 100)

0 0 0 0 1 0

(0, 4.6) (0, 4.6) (0, 4.6) (0, 4.6) (0.0, 6.2) (0, 4.6)

2 0 0 0 3 0

Data are the percentages of the compressed images being rated distinguishable from the corresponding original (noncompressed) images. Data in parentheses are the 95% confidence intervals of the percentages.

1532

Fig. 1 JPEG 2000 compression artifacts in a contrast-enhanced transverse abdominal CT image in a 58-year-old female with hepatic hemangioma. The irreversibly 5:1 compressed image (b) is indistinguishable from the original (a). At the compression level of 10:1 (c), a subtle blurring artifact appears in the central area of the hemangioma and the back muscles, which is best demonstrated if the images a and c are downloaded and displayed alternately on the same monitor. At the compression level of 15:1 (d), the manifestation of the blurring artifact is apparent and the internal texture of the hepatic hemangioma is degraded. Window width and level are 400 and 20 HU for all images

Discussion
Many researchers regard JPEG 2000 as the most sensible choice for use in the event that compression is adopted in a modern picture archiving and communication system [13]. Although there are many compression algorithms to study, our desire to produce insights that could be widely applied in the near future led us to use the JPEG 2000 algorithm. In our results, the responses by all six readers were similar and suggested that 5:1 compressed images were indistinguishable from the original images, while images compressed to a level of 10:1 or greater were distinguishable. From these results, we estimate the visually lossless threshold to be somewhere between 5:1 and 10:1 for abdominal CT images compressed using the JPEG 2000 algorithm. The effect of image compression on CT images has been studied less frequently than its effect on plain radiographs. CT images generally exhibit lower tolerance to compression than plain radiographs [5, 7, 14, 15]. The acceptable compression level was reported to be as high as 8:1 to 20:1. These reported thresholds include 8:1 for the diagnosis of acute appendicitis (using a wavelet compression algorithm) [16], 9:1 for the detection of mediastinal adenopathy and pulmonary nodules (tree-structured vector quantization) [14], 10:1 for the detection of focal hepatic lesions (a threedimensional wavelet) [7], 10:1 for nodule detection at lowdose chest CT (JPEG 2000) [17], 10:1 for the detection rate of small lung cancers at low-dose CT (JPEG and a wavelet) [18], 10:1 for the measurement of pulmonary nodule

volume at low-dose CT (JPEG 2000) [19], 20:1 for the diagnosis of coronary artery calcification (JPEG and a wavelet) [15], 20:1 for the detection of acute cerebral infarction (JPEG) [20], and 20:1 for the detection of polyps at CT colonography (a wavelet) [21]. All of these reports have concerned the evaluation of diagnostic performance (the diagnostically lossless threshold for the compression level), typically with a receiver operating characteristic study. However, image compression artifacts can be detectable even though their presence does not affect the readers performance for a given diagnostic task [2224]. The presence of such perceivable artifacts reduces the acceptance of irreversible compression among radiologists [9, 25] because a compression level tolerable to depict a gross disease may degrade fine details and texture and, therefore, possibly obscure ancillary findings [25] or induce falsepositive findings [6]. Therefore, the threshold determined by previous studies can address only narrowly defined diagnostic tasks in a given organ system [9, 20]. To provide an acceptable threshold that covers a broad range of potential abnormalities with confidence, many receiver operating characteristic studies would be required, which are timeconsuming and expensive [9]. If the compression artifact is not perceptible (i.e., if a compressed image is indistinguishable from an original image), even though the compression is irreversible, there is no basis for arguing that this visually lossless compression hinders any diagnostic accuracy [6, 8, 9]. Although the visually lossless criterion would likely allow

1533

a relatively lower compression level, this conservative criterion would be more readily acceptable even by skeptical radiologists [9]. This conception has been introduced to chest radiography by Slone et al. who concluded that JPEG compression up to 10:1 is visually lossless [6, 9]. Recently, Ringl et al. [26] reported that 3:1 (which corresponds to 4:1 in our definition) JPEG 2000 compression is visually lossless in chest CT images. Our study design was intended to be as conservative as possible in any estimate of the visually lossless threshold. We used the alternating presentation of registered images on the same monitor, because the human visual system is naturally drawn to changes in structure or brightness [9]. In addition, viewing distance was unconstrained, and images were magnified by displaying them (matrix size of 512512) on a monitor (matrix size of 1,5362,048) in one-by-one format. Results of previous studies [6, 27, 28] have shown increases in detection of compression artifacts as viewing distance decreases. We believe this research design, together with the adoption of a visually lossless threshold, should result in a very conservative and, we hope, widely accepted threshold for the compression level. Therefore, the visually lossless threshold measured in this study is the minimum (baseline) of acceptable compression level and should not be mistaken as an optimal compression level in practice. This study was conducted in the context of primary, rather than preliminary, interpretation of abdominal CT images, regardless of viewing tasks. Our results suggest that 5:1 JPEG 2000 compression is visually lossless for most readers and is, therefore, potentially acceptable for the primary interpretation of abdominal CT images without the risk of affecting diagnosis, eliminating the need to maintain the original images as the diagnostic standard. Considering the amount of modern CT data, the practical benefits of 5:1 (as a minimum) compression over reversible compression (3.5:1 in this study) are not insignificant. This reduction in data (at least 30%) would directly affect operational costs in transmission and storage, and potentially facilitate the manipulation of multiple volume datasets for comparison within a single workstation on a routine basis. These advantages would help radiologists to fully and more widely utilize the benefits of modern CT technology in spatial and temporal resolution.

The limitations of the present study are as follows. First, in order to avoid a possible clustering effect in a statistical sense, we tested only a single image per patient, which is unlike the clinical situation where radiologists scroll through a large series of images. Second, because we randomly selected the images with an intention to generalize our results throughout the abdomen, many images necessarily contained only normal structures. However, we believe that our results would be reproducible even with a study sample containing more abnormalities because our study design was sensitive enough to detect perceptible compression artifacts, regardless of the image content. Third, the tested images had fixed scanning and reconstruction parameters such as section thickness and reconstruction algorithm using a single scanner type. Although our images were likely typical of abdominal CT in most institutions, more studies are needed to determine the effects of varying imaging parameters and to further generalize our results, since the acceptable compression level can be affected by imaging parameters [5], such as the section thickness of CT images [29]. Fourth, we did not analyze the pattern(s) of compression artifacts, because this study was mainly focused on the magnitude of the artifacts. Although not stated, we have the impression that a blurring artifact mainly altering the texture of the solid organs and soft tissues made 10:1 and 15:1 compressed images distinguishable from the original images. This blurring artifact might be negligible from a diagnostic viewpoint. The denoised images at these compression levels might even be preferred from an aesthetic viewpoint [5]. However, the patterns of compression artifacts and their effects in clinical interpretation are beyond the scope of this study, and our results concerning the visually lossless threshold are still valid. In conclusion, abdominal CT images irreversibly compressed at a level of 5:1 using the JPEG 2000 algorithm are visually lossless and, therefore, potentially acceptable for primary interpretation without the risk of affecting diagnosis.
Acknowledgements This study was supported by Seoul R&BD Program, Republic of Korea (project number not assigned). We thank Jihyun Yang and Tae Ki Kim, R.T. for their assistance during image dataset preparation.

References
1. Rubin GD (2000) Data explosion: the challenge of multidetector-row CT. Eur J Radiol 36:7480 2. Rubin GD (2003) 3-D imaging with MDCT. Eur J Radiol 45(Suppl 1):S37 S41 3. Tamm EP, Thompson S, Venable SL, McEnery K (2002) Impact of multislice CT on PACS resources. J Digit Imaging 15(Suppl 1):96101 4. Lee KH, Lee HJ, Kim JH, Kang HS, Lee KW, Hong H, Chin HJ, Ha KS (2005) Managing the CT data explosion: initial experiences of archiving volumetric datasets in a mini-PACS. J Digit Imaging 18:188195 5. Erickson BJ, Manduca A, Palisson P, Persons KR, Earnest Ft, Savcenko V, Hangiandreou NJ (1998) Wavelet compression of medical images. Radiology 206:599607

1534

6. Slone RM, Muka E, Pilgram TK (2003) Irreversible JPEG compression of digital chest radiographs for primary interpretation: assessment of visually lossless threshold. Radiology 228: 425429 7. Goldberg MA, Gazelle GS, Boland GW, Hahn PF, Mayo-Smith WW, Pivovarov M, Halpern EF, Wittenberg J (1997) Focal hepatic lesions: effect of three-dimensional wavelet compression on detection at CT. Radiology 202:159165 8. Daly S (1990) Application of a noiseadaptive contrast sensitivity function to image data compression. Opt Eng 29:977987 9. Slone RM, Foos DH, Whiting BR, Muka E, Rubin DA, Pilgram TK, Kohm KS, Young SS, Ho P, Hendrickson DD (2000) Assessment of visually lossless irreversible image compression: comparison of three methods by using an image-comparison workstation. Radiology 215:543553 10. Mahesh M, Scatarige JC, Cooper J, Fishman EK (2001) Dose and pitch relationship for a particular multislice CT scanner. AJR Am J Roentgenol 177:12731275 11. Newcombe RG (1998) Two-sided confidence intervals for the single proportion: comparison of seven methods. Stat Med 17:857872 12. Liddell FD (1983) Simplified exact analysis of case-referent studies: matched pairs; dichotomous exposure. J Epidemiol Community Health 37: 8284 13. Bak PRG (2005) Does irreversible compression impact the diagnostic quality of medical images?-A review of research to date. Society for Computer Applications in Radiology Scientific Abstracts 2224

14. Cosman PC, Davidson HC, Bergin CJ, Tseng CW, Moses LE, Riskin EA, Olshen RA, Gray RM (1994) Thoracic CT images: effect of lossy image compression on diagnostic accuracy. Radiology 190:517524 15. Zheng LM, Sone S, Itani Y, Wang Q, Hanamura K, Asakura K, Li F, Yang ZG, Wang JC, Funasaka T (2000) Effect of CT digital image compression on detection of coronary artery calcification. Acta Radiol 41:116121 16. Megibow AJ, Rusinek H, Lisi V, Bennett GL, Macari M, Israel GM, Krinsky GA (2002) Computed tomography diagnosis utilizing compressed image data: an ROC analysis using acute appendicitis as a model. J Digit Imaging 15:8490 17. Ko JP, Rusinek H, Naidich DP, McGuinness G, Rubinowitz AN, Leitman BS, Martino JM (2003) Wavelet compression of low-dose chest CT data: effect on lung nodule detection. Radiology 228:7075 18. Li F, Sone S, Takashima S, Kiyono K, Yang ZG, Hasegawa M, Kawakami S, Saito A, Hanamura K, Asakura K (2001) Effects of JPEG and wavelet compression of spiral low-dose ct images on detection of small lung cancers. Acta Radiol 42:156160 19. Ko JP, Chang J, Bomsztyk E, Babb JS, Naidich DP, Rusinek H (2005) Effect of CT image compression on computerassisted lung nodule volume measurement. Radiology 237:8388 20. Ohgiya Y, Gokan T, Nobusawa H, Hirose M, Seino N, Fujisawa H, Baba M, Nagai K, Tanno K, Takeyama N, Munechika H (2003) Acute cerebral infarction: effect of JPEG compression on detection at CT. Radiology 227:124127 21. Zalis ME, Hahn PF, Arellano RS, Gazelle GS, Mueller PR (2001) CT colonography with teleradiology: effect of lossy wavelet compression on polyp detection-initial observations. Radiology 220:387392 22. MacMahon H, Doi K, Sanada S, Montner SM, Giger ML, Metz CE, Nakamori N, Yin FF, Xu XW, Yonekawa H et al (1991) Data compression: effect on diagnostic accuracy in digital chest radiography. Radiology 178:175179

23. Mori T, Nakata H (1994) Irreversible data compression in chest imaging using computed radiography: an evaluation. J Thorac Imaging 9:2330 24. Savcenko V, Erickson BJ, Palisson PM, Persons KR, Manduca A, Hartman TE, Harms GF, Brown LR (1998) Detection of subtle abnormalities on chest radiographs after irreversible compression. Radiology 206:609616 25. Kalyanpur A, Neklesa VP, Taylor CR, Daftary AR, Brink JA (2000) Evaluation of JPEG and wavelet compression of body CT images for direct digital teleradiologic transmission. Radiology 217:772779 26. Ringl H, Schernthaner RE, Bankier AA, Weber M, Prokop M, Herold CJ, Schaefer-Prokop C (2006) JPEG2000 compression of thin-section CT images of the lung: effect of compression ratio on image quality. Radiology 240:869 877 27. Cox JE, Muka E (1997) Factors affecting the selection of compression algorithms for projection radiography. Proc SPIE 3031:256264 28. Pilgram TK, Slone RM, Muka E, Cox JR, Blaine GJ (1998) Perceived fidelity of compressed and reconstructed radiological images: a preliminary exploration of compression, luminance, and viewing distance. J Digit Imaging 11:168175 29. Siddiqui KM, Siegel EL, Reiner BI, Johnson JP, Crave O, Nadar M (2004) Improved image compression at various slice thickness for multi-slice CT using 3D JPEG2000 (part 2) in comparison with conventional 2D compression. Society for Computer Applications in Radiology Scientific Abstracts 8788