| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Computer Applications |
1 From the Thoracic Division, Department of Radiology, New York University Medical Center, 560 First Ave, New York, NY 10016. From the 2001 RSNA scientific assembly. Received March 12, 2002; revision requested May 24; final revision received October 3; accepted December 14. Supported by a Scholars Award from the RSNA Research and Education Foundation. Address correspondence to J.P.K. (e-mail: jane.ko@med.nyu.edu).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
MATERIALS AND METHODS: One hundred sets of lung CT data ("cases") were compressed to 30:1, 20:1, and 10:1 levels by using a wavelet-based JPEG2000 method, resulting in 400 test cases. Each case consisted of nine 1.25-mm sections that had been obtained with 20–40 mAs. Four thoracic radiologists independently interpreted the test case images. Performance was measured by using area under the receiver operating characteristic (ROC) curve (Az) and conventional sensitivity and specificity analyses.
RESULTS: There were 51 cases with and 49 without lung nodules. Az values were 0.984, 0.988, 0.972, 0.921, respectively, for original and 10:1, 20:1, and 30:1 compressed images. Az values decreased significantly at 30:1 (P = .014) but not at 10:1 compression, with a trend toward significant decrease at 20:1 (P = .051). Specificity values were unaffected by compression (>98.0% at all compression levels). Sensitivity values were 86.3% (176 of 204 test cases with nodules), 77.9% (159 of 204 cases), 76.5% (156 of 204 cases), and 70.1% (143 of 204 cases), respectively, for original and 10:1, 20:1, and 30:1 compressed images. Results of logistic regression model analysis confirmed the significant effects of compression rate and nodule attenuation, size, and location on sensitivity (P < .05).
CONCLUSION: While no reduction in nodule detection at 10:1 compression levels was demonstrated by using ROC analysis, a significant decrease in sensitivity was identified. Further investigation is needed before widespread use of image compression technology in low-dose chest CT can be recommended.
© RSNA, 2003
Index terms: Data compression • Images, processing • Images, quality • Lung, CT, 60.12118 • Lung, nodule, 60.31
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
Image compression enables reduction of the size of image data sets, thereby increasing the speed of data transmission and decreasing data storage requirements. However, compression would be acceptable only if diagnostic accuracy and visual interpretation were not hindered. Lossless compression methods enable reversible reduction of image data without alteration, but the degree of compression is limited to 2:1. Alternatively, lossy compression methods reduce data size to levels greater than 2:1 but irreversibly change the original data.
The Joint Photographic Experts Group (JPEG) standard (3), which is a cosine transform method, and wavelet methods are the major compression methods available (3,4). Most of the more recent studies have involved the evaluation of wavelet methods, which have been shown to achieve compression levels greater than 10:1 with minimal degradation of images and were the bases for the development of a new JPEG2000 (5,6) standard for image compression (3,4).
Previous radiologic analysis of image compression techniques has primarily been concerned with the effect of lossy image compression on chest radiographs and has concentrated less on the effect of image compression on chest CT scans (7,8). To our knowledge, the effect of lossy image compression at high-spatial-resolution multi–detector row CT, particularly low-dose CT—the use of which as a method of screening for lung cancer in high-risk populations is currently being debated—has not been studied extensively.
The purpose of this study was to assess the effect of lossy JPEG2000 wavelet image compression on the detection of pulmonary nodules at low-dose CT.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
Additionally, 50 control cases were selected from CT data obtained in two patients who reportedly did not have lung nodules. Each control case comprised nine contiguous transverse sections that did not contain nodules. Each control case was selected so that the level of its sections, in terms of craniocaudal location in the thorax, was close to the level of the sections of a nodule case. Thus, CT data from a total of 25 patients—eight men and 17 women ranging in age from 39 to 75 years (mean age, 57 years)—were used.
Nodule characteristics.—Nodule characteristics—specifically, attenuation, location, and size—were recorded by the radiologist (J.P.K.) who selected the cases. Nodules were classified as solid, calcified, or subsolid in terms of attenuation. Subsolid nodules had a component of ground-glass opacity. The location of a nodule was categorized as peripheral if it was within the peripheral third of a lobe and as central if it was not. Nodules that abutted the pleura, including those adjacent to fissures, were noted. Nodule size was expressed as the largest cross-sectional dimension, and nodule measurement was performed by using electronic calipers at an image workstation (Wizard; Siemens Medical Systems, Iselin, NJ). Nodules were subcategorized as those equal to or smaller than 5 mm and those larger than 5 mm. More than one nodule appeared on the center section in 11 cases. The coordinates of all nodules were recorded.
Imaging protocol.—Low-dose screening chest CT had been performed with a multi–detector row CT scanner with an adaptive array detector (Somatom Volume Zoom 4; Siemens Medical Systems). All CT images had been acquired helically in one breath hold without the administration of intravenous contrast material and with collimation of 1.0 mm, tube current of 20–40 mAs, and 120 kV. Pitch was variable and ranged from 1.4 to 1.7. CT data were reconstructed into 1.25-mm transverse sections at 1.0-mm intervals; a high-frequency reconstruction algorithm and a matrix of 512 x 512 pixels were used. The field of view ranged between 28 and 35 cm.
Wavelet Compression of CT Data
All identifying patient information was removed from the images. CT data in Digital Imaging and Communications in Medicine (DICOM) version 3.0 format were transferred within the hospital computer network to a research computer (Dell Dimension, Austin, Tex) with a Windows NT operating system. A Java implementation (JJ2000; available for noncommercial use at jj2000.epfl.ch) of the JPEG2000 standard was adapted to 16-bit medical images. JPEG2000 is a discrete wavelet transform that uses a Daubechies biorthogonal basis. JPEG2000 supports error resilience, random access, and the ability to perform simple manipulation of compressed-domain data.
Each image of the 100 original cases was compressed to levels of 10:1, 20:1, and 30:1, yielding 400 test cases for interpretation. The actual mean compression levels achieved for the 10:1, 20:1, and 30:1 nominal levels, respectively, were 10.00 ± 0.04 (SD), 19.97 ± 0.10, and 30.05 ± 0.19.
Image Interpretation
Interpretation of images was performed at a personal computer (DeskPro; Compaq, Houston, Tex) with a Windows NT operating system and a diagnostic-quality picture archiving and communication systems (PACS) monitor (M21LMAX; Image Systems, Minnetonka, Minn). The monitor satisfied the American College of Radiology standards for teleradiology and digital image data management (9,10) and had a luminance of 65 foot-lamberts, an aperture grille pitch of 0.25 mm, a resolution of 1,200 x 1,600, and a refresh rate of 75-Hz.
Four thoracic radiologists, who had 2–22 years of experience in chest radiology, served as independent readers. The radiologist who selected the cases did not serve as a reader so that memory bias could be minimized. Each reader evaluated the 400 test cases in four 1
-hour sessions that were separated by at least 7 days. During each session a reader evaluated 100 test cases, with a short break in the middle of the session to minimize fatigue. Test cases in reading sessions 1, 2, and 3 consisted of images compressed to 30:1, 20:1, and 10:1, respectively, and at session 4 original images were interpreted. Within each session, images from the nodule and control test cases were randomly presented to the readers. The randomization scheme ensured that nodule images that differed in image compression rates were presented and interpreted after a sufficient time delay. The readers were unaware of the ratio of nodule test cases to control test cases, the levels of compression being evaluated, and the order in which images compressed with the different levels were presented. Images were presented with specific window settings (window level, -600 HU; window width, 1,500 HU), but each reader was able to adjust window settings to his or her preference.
For each test case, the readers assessed whether one or more nodules were present specifically on the fifth of nine contiguous sections. For each nodule they identified, the readers rated their degree of confidence according to a scoring system (Table 1) and marked the coordinates of the nodule.
|
Statistical Analysis
Reader performance was expressed as the area under the receiver operating characteristic (ROC) curve (Az) (11). Statistically significant differences between compression levels were tested by using a univariate z-score test area. A statistical software package (SSPS, Chicago, Ill) was used. Differences in the Az values between compression levels were identified by using a one-tailed Student t test. Conventional measurements of accuracy, sensitivity, and specificity were calculated by using ratings of 1 and 2 as negative responses and 3 and 4 as positive responses; significant differences were identified by using one-tailed 
2 tests. P < .05 was considered to indicate a statistically significant difference.
Logistic regression analysis was used to identify factors predictive of a true-positive reading. In this analysis, the outcome variable was the binary assessment (nodule detected or not detected) for each positive case, while the potentially predictive factors included in the model were compression level (coded as a continuous variable), nodule opacity (ground glass vs solid), nodule size category (≤5 mm, >5 mm), and nodule location. Nodule location was represented in the model as two binary variables: (a) central location and (b) adjacency to pleura. After the main effect was determined, the interactions between the four variables describing nodule characteristics and the compression levels were determined.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
The sizes of the nodules evaluated ranged between 2 and 15 mm in diameter (mean, 5.8 mm ± 3.2 [SD]; median, 5 mm). Among the positive cases, 31 had nodules in the right lung and 20 had nodules in the left lung. There were 15 subsolid nodules, 13 of which had pure ground-glass opacity and two of which had mixed ground-glass and solid opacity. Thirty-three nodules were solid, and three were calcified. There were 32 nodules that abutted the pleura, with 24 touching the costal pleura, seven adjacent to a fissure, and one touching both the costal pleura and fissure. Seven nodules were central, and 44 were peripheral in location.
The mean Az values for all four radiologists were 0.984 ± 0.012 (SD), 0.988 ± 0.007, 0.972 ± 0.014, and 0.921 ± 0.034, respectively, for the original images and the images compressed at 10:1, 20:1, and 30:1 (Figs 1–3). Compared with Az values for interpretation of uncompressed images, the Az values decreased significantly (P = .014) when readers interpreted images at 30:1 compression. There was a trend toward a significant difference at 20:1 compression (P = .051), and no significant difference was identified at 10:1 compression. Az values for all four radiologists decreased as compression levels increased from 10:1 to 30:1 (Table 2). Statistically significant differences between the correlated ROC curves for images at 10:1 compression and images at 30:1 compression were found for three of four readers with the univariate z-score area test. The SDs of the mean Az values for all readers increased as compression levels increased, indicating increased variation in performance (Table 2).
|
|
|
|
2 test) versus noncompressed images. Specificity was very high and was unaffected (P > .05) by image compression; specificity values were 98.5% (193 of 196 test cases), 99.0% (194 of 196 test cases), 99.0% (195 of 196 test cases), and 98.5% (193 of 196 test cases), respectively, for original images and 10:1, 20:1, and 30:1 compressed images. Linear regression model analysis revealed a significant main effect for the compression rate as well as for each nodule characteristic, meaning that each variable had an independent effect on sensitivity (Table 3). The ORs and 95% CIs for each characteristic are reported in Table 3. The interactions between nodule characteristics and compression rate were significant for nodule location and size (Table 3).
|
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
There is preliminary evidence that image compression can actually improve image quality by reducing image noise, but at high levels of compression, texture and fine details can be lost, with subsequent loss of image quality (12–14). Two-dimensional discrete cosine transform and wavelet methods are most commonly used for image compression, although three-dimensional wavelet compression (15–17) and raw data compression (18) have also been studied. A group of wavelet methods that are based on a discrete wavelet transform that uses Daubechies biorthogonal basis are currently being developed as a JPEG2000 standard for general-purpose image compression.
Previous evaluation of radiologic image compression techniques has primarily focused on their use in chest radiography. In chest radiography, direct cosine transform methods (12,19–23) preserved image quality and were described as "visually lossless" at levels up to 11:1 (12) and 13:1 (22). There was also no significant decrease in accuracy at compression levels of 20–25:1 for diagnosis of nodules (19,20,22) and interstitial lung disease (20,23) on chest radiographs. Wavelet compression at chest radiography (13,14,24) preserved image quality at levels up to 11:1 (12), 20:1 (24), and 40:1 (13) and diagnostic accuracy at levels up to 80:1 (14). The effect of image compression at CT (7,8,17,25,26), particularly chest CT (7,8), has been studied less frequently than its effect at chest radiology. CT images, particularly those reconstructed by using a high-frequency kernel, may be less amenable to compression than chest radiographs because of the lower percentage of energy in the lowest frequency subband at CT (4).
In our study, which involved 100 cases and four radiologists, we did not find a reduction in Az values for nodule detection when we evaluated uncompressed and 10:1 compressed images. However, we did observe significant decreases in sensitivity—from 86.3% with original images to 77.9% with 10:1 and 76.5% with 20:1 compressed images. We did not observe any change in specificity due to compression, meaning that the JPEG2000 algorithm did not introduce artifacts that were construed as nodules. Although our study results corroborate with those of Li et al (8), it should be noted that there were major differences in the design of the two studies, including our use of high-spatial-resolution sections, multiple contiguous sections per case, a diagnostic-quality PACS monitor, and an experimental design that minimized reader memory bias.
Results of logistic regression analysis in our study revealed that compression rate and nodule attenuation, size, and location had independent and significant effects on sensitivity. The nodule characteristics that had the strongest negative effects on sensitivity were size of 5 mm or less (OR, 0.27; 95% CI, 0.18, 0.42) and ground-glass opacity (OR, 0.31; 95% CI, 0.19, 0.49). Rusinek et al (27) reported sensitivity values of 58%–62% for detection of central nodules—versus 74% for detection of peripheral nodules—at standard and low-dose chest CT. Our study also revealed higher detection rates for nodules that touch the pleura. The higher sensitivity values for nodules in the peripheral and, particularly, the subpleural regions may be related to the paucity of vessels in these regions.
Our results suggest that the loss of sensitivity with increasing image compression levels is related to small nodule size and central nodule location. Although compression may result in a change in the appearance of the texture of ground-glass nodules and potentially decrease detection rates, results of our logistic regression analysis did not reveal a significant interaction between nodule attenuation and image compression level.
We concentrated on the evaluation of low-dose CT images with the expectation that our results would be applicable to common diagnostic CT techniques used to evaluate individuals suspected or known to have pulmonary nodules. According to the findings of Rusinek et al (27), nodule detection is decreased at low-dose chest CT. Diagnostic CT images may be affected by image compression to a lesser degree than low-dose CT images; this may be a topic for future investigation. Before image compression at CT performed for lung nodule detection is widely accepted, we need to ensure that use of image compression does not compromise detection of other pulmonary diseases that may be coexistent, such as infiltrative lung disease.
In conclusion, results of ROC analysis did not demonstrate a reduction in diagnostic accuracy at 10:1 levels of JPEG2000 wavelet compression. However, a decrease in sensitivity at this level was identified, highlighting the need for further investigation before image compression technology is widely used at low-dose chest CT.
|
ACKNOWLEDGMENTS |
|---|
|
FOOTNOTES |
|---|
Author contributions: Guarantors of integrity of entire study, J.P.K., H.R.; study concepts and design, J.P.K., H.R.; literature research, J.P.K.; clinical studies, D.P.N., G.M., A.N.R., B.S.L., J.M.M.; data acquisition, J.P.K.; data analysis/interpretation, J.P.K., H.R.; statistical analysis, H.R.; manuscript preparation, J.P.K., H.R.; manuscript definition of intellectual content, J.P.K., H.R., D.P.N.; manuscript editing, revision/review, and final version approval, all authors.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
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]
|
|
|
|
 |
|
 A. Fidler, B. Likar, and H. Rupert Ringl What Is Wrong with Compression Ratio in Lossy Image Compression? Radiology, October 1, 2007; 245(1): 299 - 300. [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]
|
|
|
|
|
|
Y. J. Jeong, C. A. Yi, and K. S. Lee Solitary Pulmonary Nodules: Detection, Characterization, and Guidance for Further Diagnostic Workup and Treatment Am. J. Roentgenol., January 1, 2007; 188(1): 57 - 68. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
H. Ringl, R. E. Schernthaner, A. A. Bankier, M. Weber, M. Prokop, C. J. Herold, and C. Schaefer-Prokop JPEG2000 Compression of Thin-Section CT Images of the Lung: Effect of Compression Ratio on Image Quality Radiology, September 1, 2006; 240(3): 869 - 877. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
J. P. Ko, J. Chang, E. Bomsztyk, J. S. Babb, D. P. Naidich, and H. Rusinek Effect of CT Image Compression on Computer-assisted Lung Nodule Volume Measurement Radiology, October 1, 2005; 237(1): 83 - 88. [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]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
