HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS		QUICK SEARCH:   [advanced] Author: Keyword(s): Year:  Vol:  Page:

This Article Abstract Figures Only Full Text (PDF) Submit a response Alert me when this article is cited Alert me when eLetters are posted Alert me if a correction is posted Citation Map Services Email this article to a friend Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Slone, R. M. Articles by Pilgram, T. K. Search for Related Content PubMed PubMed Citation Articles by Slone, R. M. Articles by Pilgram, T. K.

Irreversible JPEG Compression of Digital Chest Radiographs for Primary Interpretation: Assessment of Visually Lossless Threshold¹

¹ From the Electronic Radiology Laboratory, Mallinckrodt Institute of Radiology, Washington University School of Medicine, Box 8131, 510 S Kingshighway Blvd, St Louis, MO 63110. Received December 6, 2001; revision requested February 20, 2002; final revision received October 28; accepted November 11. Address correspondence to E.M. (e-mail: ete@wuerl.wustl.edu).

	ABSTRACT

TOP ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION REFERENCES

 
PURPOSE: To determine if digital chest images could be compressed in a primary interpretation context without perceived loss of fidelity (below the visually lossless threshold) at transilluminated film or cathode ray tube (CRT) display.

MATERIALS AND METHODS: One hundred forty-four posteroanterior radiographs were obtained with a digital chest radiography system. At both film and CRT display, an identified original image was presented side by side with a replicate, which was either an unaltered image or an image that had been Joint Photographic Experts Group (JPEG) compressed to 10:1, 20:1, or 50:1 and reconstructed. Each of the 10 readers indicated whether the replicate was "indistinguishable from the original" or "degraded" at clinical reading distance and at close inspection. The readers’ ability to detect compressed images was examined for patterns; 95% CIs were used for statistical testing.

RESULTS: With transilluminated film at clinical reading distance, readers were as likely to rate originals (48 [20%] of 240 readings) as degraded as they were to rate 20:1 replicates (106 [22%] of 480 readings) as degraded, but they frequently identified 50:1 replicates (283 [59%] of 480 readings) as degraded. At close inspection, 20:1 replicates (163 [34%] of 480 readings) were often identified as degraded, but 10:1 replicates (19 [8%] of 240 readings) were not identified as degraded more often than originals (17 [7%] of 240 readings). With CRT display, the results were nearly identical.

CONCLUSION: At reading distance for primary interpretation, full-size digital chest radiographs that have been JPEG compressed to 10:1 or 20:1 and reconstructed are visually lossless at film or CRT display. Images compressed to 10:1 remain visually lossless at close inspection.

© RSNA, 2003

Index terms: Images, digitization • Images, display • Images, processing • Images, quality • Radiography, digital

	INTRODUCTION

TOP ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION REFERENCES

 
Image compression can greatly reduce data transfer rates and storage requirements and will certainly become an integral component of electronic imaging systems. A level of compression great enough to result in a meaningful reduction in image data size mathematically prevents complete restoration of the original, a situation known as irreversible compression. Accordingly, there is understandable reluctance on the part of standards groups to specify, and professional organizations to sanction, acceptable levels of irreversible compression for primary interpretation of radiographic images.

However, whether mathematically irreversible compression—so-called lossy compression—leads to perceptible degradation in image fidelity is unclear. The acceptable degree of compression is dependent on several independent parameters, including those involved in the image source (modality, digitization, spatial resolution, and noise sources), image processing (data space, tonal characteristics, frequency enhancement, and display corrections), display (luminance, frequency response, and noise), the human visual system, and reading tasks (reading conditions, reader experience, specific task, and psychophysical performance issues). Therefore, lossy image compression warrants careful investigation.

Our principal hypothesis is that mathematical loss can exist before observable image degradation occurs. This amount of lossy compression is termed visually lossless (1) because the compressed and reconstructed image appears to be indistinguishable from the original image. The purpose of our study was to determine if digital chest images could be compressed in a primary interpretation context without perceived loss of fidelity (below the visually lossless threshold [VLT]) at transilluminated film or cathode ray tube (CRT) display.

We have chosen a visually based image comparison study rather than an interpretative or diagnostic study of a single image because we believe radiologists are extremely sensitive to visual differences between images even though they are able to make accurate diagnoses from images of varying quality.

	MATERIALS AND METHODS

TOP ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION REFERENCES

 
Overview
We used a modified forced-choice study to evaluate loss of fidelity in compressed and reconstructed images—herein called replicates—intended for primary interpretation. The study involved 10 readers and 144 patient images—digital posteroanterior chest radiographs, which included normal, abnormal, or incidental findings. The replicates included 24 unaltered images and 120 compressed images prepared by using public domain Joint Photographic Experts Group (JPEG) compression software (2,3). For this experiment, we chose compression levels that our experience suggested would be below the VLT (JPEG compression level 10) for 24 images, at the VLT (JPEG compression level 20) for 48 images, and above the VLT (JPEG compression level 50) for 48 images.

Reading was conducted with 14 x 17-inch transilluminated images and with high-performance CRT displays. Each original image, which was identified as such, was presented side by side with a randomized replicate, and the reader was asked to indicate if the replicate was indistinguishable from the original or if detectable degradation or an artifact was present. Readers first performed the image comparison at a clinical reading distance and then performed it again at an unrestrained close inspection distance. Reader response and confidence ratings were recorded, as was the actual reading distance used.

Images
Our use of clinical images was approved by the Internal Review Board for Human Studies of the hospital and the institutional review board of the radiology department with which we are affiliated. Informed consent was not required. Over an 8-day period, 144 consecutive digital posteroanterior chest radiographs were acquired in the outpatient and admitting area of our hospital with a commercial selenium detector system (Thoravision; Philips Medical Systems, Shelton, Conn) and were recorded on optical disk. These patient images were obtained in approximately even numbers of men (n = 76) and women (n = 68). To obtain additional information about each patient image, a radiologist (R.M.S.) read the accompanying official report. The images were classified into one of three categories—as showing normal (25 of 144 images), abnormal (70 of 144 images), or incidental (49 of 144 images) findings; Table 1 provides details of the latter two categories.

As shown in Table 2, the 144 images were block randomized into one of four compression categories (unaltered original images and images processed at one of three compression levels). Because compression artifact is readily apparent near large signal transitions, we noted for later data analysis the presence of hardware, including implantable devices and metallic material such as pacemakers, defibrillators, clips, sternal wires, central venous catheters, surgical staples, and vena caval filters.

Each image was received in digital form as 15 bits/pixel packed into 2 bytes; this constitutes the bit depth (16 bits) of the original image. Because image data represent a relative log₁₀ luminance range of 3.0 scaled by 10,000, each pixel was normalized by the ratio 4,096:30,000 to comply with the 12-bits/pixel JPEG algorithm. Except for those images where an unaltered image was used as a control, each image was compressed at the particular compression prescribed in the block randomization plan. Except for initial data input and output, all calculations, including those associated with the JPEG algorithm, were completed in floating point form. After reconstruction, each replicate and unaltered original was prepared for display with film and CRT by using existing hardware and software (described in detail below) from our laboratory.

Image Presentation and Scoring
Because we wished to be as conservative as possible in any estimate of the VLT, we decided to use a modified forced-choice method in which the original image was identified for the reader. While some may challenge the use of this approach versus a true two-alternative forced-choice method, it seemed to us that knowledge of which image was the original would bias the reader, in a conservative manner, to favor the original image.

With a random number sequence, the 144 image pairs were presented in a side-by-side orientation. The identified original was always placed to the left-hand side, and the replicate was always placed to the right-hand side. Each reading session, at which all images were presented entirely with one of the two viewing modalities, was supervised by a research assistant and conducted at the convenience of the reader, without time constraint. Each session was divided into two parts according to viewing distance, as explained below. Room lights were dimmed enough to produce the desired luminance dynamic range for the display being used.

The readers were asked to perform according to the following instructions:

1. The image on the left is an unaltered original digital chest radiograph. The image on the right is either an unaltered original or an image that has been compressed to one of several possible levels and then reconstructed. Please choose one of the following responses: O = the image on the right is indistinguishable in any way from the original or D = there is visual evidence of some degradation in quality, loss of fidelity, or presence of an artifact in the image on the right.

2. Please indicate your level of confidence in your response by using the following scale: 1 = uncertain, 2 = somewhat confident, 3 = confident, 4 = very confident, and 5 = certain.

During the first part of a session, readers were asked to limit the reading distance to that which would be used for clinical work, including image screening and inspection of detected abnormalities. This clinical distance was measured and recorded by the research assistant. A second reading of the same image pair was then completed as the second part of the session, and in this part the reader was permitted (and encouraged) to inspect the images as closely as needed to discern visual differences. This inspection distance was measured and recorded by the research assistant. The film reading was performed first by five readers who then, after a sufficient delay, performed the CRT reading. The other five readers performed the CRT reading first and the film reading second. The mean time between the readers’ completion of the film readings and beginning of the CRT readings was 145 days (range, 126–174 days).

Readers
Ten readers were selected from our diagnostic radiology program. They included two junior residents (1–2 years experience each), two senior residents (more than 3 years experience each), two general radiology faculty members, and four chest radiology faculty members; each faculty member had more than 5 years experience beyond training. All 10 readers read all of the image combinations.

Displays
The reading experiment was conducted with 14 x 17-inch transilluminated film produced with a laser printer-processor system (Eastman Kodak, Rochester, NY) and with a monochrome (P-45 phosphor), high-spatial-resolution (2,000 x 2,500 pixels), 21-inch CRT (Data Ray, Westminster, Colo) with matching video hardware (Dome Imaging, Waltham, Mass). The CRT software presented the image with a style that we call the "film look," wherein the relative luminance distribution of the image on the CRT display matches that on film. Use of this style results in images that appear to the readers to be essentially identical with both displays.

The maximum luminance of the film light box was approximately 2,400 candelas (cd) per square meter, and the maximum luminance of the CRT was approximately 225 cd/m²; both display modalities had a log₁₀ luminance dynamic range of 2.8. Laser printer and processor calibration was maintained during the production of the 288 film prints. CRT luminance calibration was routinely checked and maintained during the reading study. Because it was known that the spatial frequency response of the CRT was limited in comparison with that of transilluminated film, a precompensated CRT version was generated by using the common unsharp mask algorithm. On the basis of several thoracic radiologists’ preference, the unsharp mask parameters were as follows: kernel of 5 and boost of 0.2 (gain of 1.2).

Statistical Analysis
We primarily performed descriptive statistical analysis so that we could discover and present clear trends in the data. The percentage of images rated as indistinguishable from the original was calculated for each compression level. When that percentage was the same as that for uncompressed images, we believed that the compressed images were indistinguishable from the originals. To aid in this comparison, we calculated 95% CIs (4). If one percentage lay outside the 95% CIs of another, we considered the difference to be statistically significant when P < .05. We used the number of images as the sample size in the calculations of confidence. Because the performance of the 10 readers showed very high correlations and multiple CIs were calculated, this seemed like a simple conservative approach. Bubble plots—in which the area of the bubble is proportional to the number of data points with identical values—were used to display the results of individual readers.

	RESULTS

TOP ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION REFERENCES

 
When all reader data were combined, a fairly clear pattern emerged; however, readers differed in their ability to detect compressed and reconstructed replicates. At the clinical reading distance (Figs 1, 2), readers as a group were unable to detect 10:1 replicates or 20:1 replicates and were able to detect 50:1 replicates only about half the time. At close inspection distance (Figs 3, 4), readers as a group could not detect 10:1 replicates, could sometimes detect 20:1 replicates, and could usually detect 50:1 replicates. These results held true with either film or CRT display and are summarized in Table 4.

View larger version (34K): [in this window] [in a new window] [Download PPT slide]   Figure 1. Bar graph shows the percentage of replicates at each compression level rated as indistinguishable from the original at clinical reading distance (results are averaged for all readers). The gray bars represent film readings, and the white bars represent CRT readings. Error bars denote 95% CIs based on the number of images (24 images at 1:1 compression, 24 at 10:1, 48 at 20:1, and 48 at 50:1).

View larger version (41K): [in this window] [in a new window] [Download PPT slide]   Figure 2. Bubble plot shows the distribution of the percentage of replicates at each compression level rated as indistinguishable from the original at clinical reading distance by each reader. The gray bubbles represent film readings, and the white bubbles represent CRT readings. Bubble area is proportional to the number of data points, with the smallest bubble representing one reader.

View larger version (32K): [in this window] [in a new window] [Download PPT slide]   Figure 3. Bar graph shows the percentage of replicates at each compression level rated as indistinguishable from the original at close inspection distance (results are averaged for all readers). The gray bars represent film readings, and the white bars represent CRT readings. Error bars denote 95% CIs based on the number of images (24 images at 1:1 compression, 24 at 10:1, 48 at 20:1, and 48 at 50:1).

View larger version (36K): [in this window] [in a new window] [Download PPT slide]   Figure 4. Bubble plot shows the distribution of the percentage of replicates at each compression level rated as indistinguishable from the original at close inspection distance by each reader. The gray bubbles represent film readings, and the white bubbles represent CRT readings. Bubble area is proportional to the number of data points, with the smallest bubble representing one reader.

Reading Distance
Because the size of the image was essentially the same at both film and CRT display, we present combined results for reading distance. There was no tendency for distance to vary with the diagnostic content of the image (normal, abnormal, or incidental findings), the presence of hardware, or the readers’ confidence in their judgment of image quality. There was a great deal of individual variation, but the average close inspection distance for all readers was about half the clinical reading distance. Individual mean clinical reading distance ranged from 37 to 57 cm, and individual mean inspection distance ranged from 14 to 33 cm (P < .001, analysis of variance). The presumed difference in vision among the readers, which we did not assess, prevents us from drawing any causal conclusion from these results.

Image Content
There was no association between reader accuracy and the diagnostic content of the image (normal, abnormal, or incidental findings) or the presence of any hardware.

	DISCUSSION

TOP ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION REFERENCES

 
Published reports of lossy image compression (5–11) have typically concerned the evaluation of diagnostic performance with a receiver operating characteristic study. Studies of this type require an extensive image collection process and a substantial amount of reader time. It is likely that in our study, readily observable visual distortion was present at the higher compression levels used, as it should be in a properly designed study. However, visual distortion is disturbing to some radiologists; it increases their apprehension and may result in "overcalling pathology." We decided to conduct a more conservative study in which a visually rather than a diagnostically lossless threshold was identified because it is our belief that a visually lossless compressed and reconstructed image is also diagnostically lossless.

We recognize the importance of CRT display units in an electronic radiology system. However, CRT display is not physically equivalent to transilluminated film display with a light box, owing to properties of the human visual system (12,13) and the CRT (14). Accordingly, there exists a risk that the performance of a radiologist is affected when he or she reads original images with CRT display. Additionally, tolerable compression levels might be higher at CRT display than at transilluminated film display. Stated from a different perspective, at equal compression ratios, visible artifacts are likely to be more apparent on transilluminated film. Our goal, then, was to complete a study of lossy image compression that included both film and CRT displays in an attempt to be conservative in our assessment of the VLT.

Finally, although there are many compression algorithms to study, our desire to produce insights that could be widely applied led us to use the JPEG algorithm. In considering image compression at the VLT, we determined in our pilot studies that JPEG performed at least as well as the currently popular wavelet-based algorithm (15).

Our decision to use posteroanterior chest radiographs was based on the high frequency of acquisition of these radiographs in a large hospital setting. However, one may question the broad applicability of our conclusions given this unique image. Our experience suggests that readers are critically sensitive to distortions in image areas, such as the shoulder areas of a posteroanterior chest radiograph, that show small luminance changes over large distances. Similar regions appear in other projection radiographs such as lateral views of the cervical spine and views of dense breasts. Thus, we feel that our results apply more broadly.

On the basis of our results with full-size digital posteroanterior chest images read with film and CRT display, we conclude that, at clinical reading distances, 20:1 JPEG compressed (0.80 bits/pixel) and reconstructed images are visually lossless, while at close inspection only 10:1 JPEG compressed (1.6 bits/pixel) and reconstructed images are visually lossless.

The practical benefits of 10:1 compression are important. A 10-fold reduction in data transmission time and storage volume would directly affect operational costs. Although it is difficult to quantify, there would be a productivity enhancement in the interpretation of images if current and previous images (assuming the previous images were electronically stored) could be compared on a routine basis without substantial delay. We believe that radiologists will accept such an imaging system as reasonable and begin to use it for primary interpretation.

	FOOTNOTES

 
Abbreviations: CRT = cathode ray tube, JPEG = Joint Photographic Experts Group, VLT = visually lossless threshold

Author contributions: Guarantors of integrity of entire study, R.M.S., E.M.; study concepts, R.M.S., E.M.; study design, all authors; literature research, R.M.S., E.M.; clinical studies, all authors; data acquisition, R.M.S., E.M.; data analysis/interpretation, all authors; statistical analysis, T.K.P.; manuscript preparation, definition of intellectual content, editing, revision/review, and final version approval, all authors

	REFERENCES

TOP ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION REFERENCES

 

1. Daly S. Application of noise-adaptive contrast sensitivity function to image data compression. Opt Eng 1990; 29:977-987.[CrossRef]
2. Pennebaker WB, Mitchell JL. JPEG still image data compression standard New York, NY: Van Nostrand Reinhold, 1993.
3. Independent JPEG Group home page. Available at: www.ijg.org. Accessed June 4 2003.
4. Fleiss JL. Statistical methods for rates and proportions New York, NY: Wiley, 1981.
5. Ishigaki T, Sakuma S, Ikeda M, Itoh Y, Suzuki M, Iwai S. Clinical evaluation of irreversible image compression: analysis of chest imaging with computed radiography. Radiology 1990; 175:739-743.[Abstract/Free Full Text]
6. MacMahon H, Doi K, Sanada S, et al. Data compression: effect on diagnostic accuracy in digital chest radiography. Radiology 1991; 178:175-179.[Abstract/Free Full Text]
7. Aberle DR, Gleeson F, Sayre JW, et al. The effect of irreversible image compression on diagnostic accuracy in thoracic imaging. Invest Radiol 1993; 28:398-403.[CrossRef][Medline]
8. Collins CA, Lane D, Frank M, et al. Design of a receiver operating characteristic (ROC) study of 10:1 lossy image compression. Proc SPIE 1994; 2166:149-158.[CrossRef]
9. Mori T, Nakata H. Irreversible data compression in chest imaging using computed radiography: an evaluation. J Thorac Imaging 1994; 9:23-30.[Medline]
10. Kido S, Ikezoe J, Kondoh H, et al. Detection of subtle interstitial abnormalities of the lungs on digitized chest radiographs: acceptable data compression ratios. AJR Am J Roentgenol 1996; 167:111-115.[Abstract/Free Full Text]
11. Uchida K, Nakamura K, Watanabe H, et al. Clinical evaluation of irreversible data compression for computed radiography in excretory urography. J Digit Imaging 1996; 9:145-149.[Medline]
12. Pilgram TK, Slone RM, Muka E, Cox JR, Blaine GJ. Perceived fidelity of compressed and reconstructed radiological images: a preliminary exploration of compression, luminance and viewing distance. J Digit Imaging 1996; 11:168-175.
13. Blume H, Daly S, Muka E. Presentation of medical images on CRT displays: a renewed proposal for a display function standard. Proc SPIE 1993; 1897:215-231.[CrossRef]
14. Muka E, Mertelmeier T, Slone RM, Senol E. Impact of phosphor luminance noise on the specification of high-resolution CRT displays for medical imaging. Proc SPIE 1997; 3031:1-12.
15. Foos DH, Muka E, Slone RM, et al. JPEG 2000 compression of medical imagery. Proc SPIE 2000; 3980:85-96.[CrossRef]

This article has been cited by other articles:

				K. J. Kim, B. Kim, K. H. Lee, T. J. Kim, R. Mantiuk, H.-S. Kang, and Y. H. Kim Regional Difference in Compression Artifacts in Low-Dose Chest CT Images: Effects of Mathematical and Perceptual Factors Am. J. Roentgenol., August 1, 2008; 191(2): W30 - W37. [Abstract] [Full Text] [PDF]

				V. Bajpai, K. H. Lee, B. Kim, K. J. Kim, T. J. Kim, Y. H. Kim, and H. S. Kang Differences in Compression Artifacts on Thin- and Thick-Section Lung CT Images Am. J. Roentgenol., August 1, 2008; 191(2): W38 - W43. [Abstract] [Full Text] [PDF]

				B. Kim, K. H. Lee, K. J. Kim, R. Mantiuk, H.-r. Kim, and Y. H. Kim Artifacts in Slab Average-Intensity-Projection Images Reformatted from JPEG 2000 Compressed Thin-Section Abdominal CT Data Sets Am. J. Roentgenol., June 1, 2008; 190(6): W342 - W350. [Abstract] [Full Text] [PDF]

				B. Kim, K. H. Lee, K. J. Kim, R. Mantiuk, S. Hahn, T. J. Kim, and Y. H. Kim Prediction of Perceptible Artifacts in JPEG 2000 Compressed Chest CT Images Using Mathematical and Perceptual Quality Metrics Am. J. Roentgenol., February 1, 2008; 190(2): 328 - 334. [Abstract] [Full Text] [PDF]

				H. S. Woo, K. J. Kim, T. J. Kim, S. Hahn, B. Kim, Y. H. Kim, and K. H. Lee JPEG 2000 Compression of Abdominal CT: Difference in Tolerance Between Thin- and Thick-Section Images Am. J. Roentgenol., September 1, 2007; 189(3): 535 - 541. [Abstract] [Full Text] [PDF]

				S. Suryanarayanan, A. Karellas, S. Vedantham, S. M. Waldrop, and C. J. D'Orsi Detection of Simulated Lesions on Data-compressed Digital Mammograms Radiology, July 1, 2005; 236(1): 31 - 36. [Abstract] [Full Text] [PDF]

This Article Abstract Figures Only Full Text (PDF) Submit a response Alert me when this article is cited Alert me when eLetters are posted Alert me if a correction is posted Citation Map Services Email this article to a friend Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Slone, R. M. Articles by Pilgram, T. K. Search for Related Content PubMed PubMed Citation Articles by Slone, R. M. Articles by Pilgram, T. K.