| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Original Research |
1 Department of Radiology, Seoul National University Bundang Hospital, 300
Gumi-dong, Bundang-gu, Seongnam-si, Gyeonggi-do, Seoul 463-707, Korea.
2 Institute of Radiation Medicine, Seoul National University Medical Research
Center, Seoul, Korea.
3 Medical Research Collaborating Center, Seoul National University Hospital,
Seoul National University College of Medicine, Seoul, Korea.
Received November 21, 2006;
accepted after revision March 26, 2007.
Address correspondence to K. H. Lee
(kholee{at}snubhrad.snu.ac.kr).
Abstract
 Top Abstract IntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
MATERIALS AND METHODS. One hundred 0.67-mm-thick and corresponding 5-mm-thick images were compressed to four different levels: reversible and irreversible 6:1, 10:1, and 15:1. Five radiologists determined if the compressed images were distinguishable from the originals. The percentage of distinguishable pairs and peak signal-to-noise ratio (PSNR) were compared between the thin and thick sections. The visually lossless threshold was estimated by comparing the percentages of the distinguishable pairs between each irreversible compression and the reversible compression. Paired Student's t tests and exact tests for paired proportions were used.
RESULTS. Thin sections had smaller PSNRs at each compression level (p < 0.001). According to the pooled responses, the percentages of distinguishable pairs for the thin and thick sections, respectively, were 0% (0/100) and 0% at reversible compression, 27% and 0% at 6:1 (p < 0.001), 100% and 80% at 10:1 (p < 0.001), and 100% and 100% at 15:1. Artifacts increased significantly (p < 0.001) at 6:1 or more for the thin sections and at 10:1 and 15:1 for the thick sections, indicating that the visually lossless thresholds were below 6:1 and between 6:1 and 10:1, respectively.
CONCLUSION. Thin-section abdominal CT images are less tolerant of compression, and a lower compression level should be used for the visually lossless threshold.
Keywords: CT • data compression • JPEG 2000 • visually lossless threshold
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
The acceptable compression level is dependent on several independent parameters [6, 16], including compression algorithm, reviewing task, image content, and imaging parameters such as section thickness of CT images [17, 18]. For instance, it has been reported that detection performance for liver nodules in 5-mm-thick images is preserved with up to 10:1 compression [8]; however, it is uncertain if this compression level is acceptable for submillimeter-thick images generated from the newer scanners.
The purpose of this study is to show the difference in tolerance (defined as the range of compression at which the decompressed image is acceptable) [19] to the Joint Photographic Experts Group (JPEG) 2000 compression between thin- and thick-section abdominal CT images. Of the many CT parameters, we have focused on the section thickness because we believe it is the most distinguishing feature of modern CT scanners over previous scanners.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
CT
This study included 100 consecutive adult patients (61 males and 39
females; age range, 16-95 years) who underwent single-phase contrast-enhanced
abdominal CT with a 64-MDCT scanner (Brilliance, Philips Medical Systems)
during a period of 10 days in February 2006. We did not ascertain the reasons
for CT examination in these patients.
IV nonionic contrast material (2 mL/kg iopromide [Ultravist 370, Schering]) was administered at a rate of 3 mL/s. Bolus-tracking software was used to trigger scanning 60 seconds after the aortic enhancement reached a 150-H threshold. Raw projection data were obtained using the following scanning parameters: scanning range, from diaphragm to symphysis pubis; detector collimation, 0.625 mm; gantry rotation time, 0.42 second; tube potential, 120 kVp; and pitch, 1.11-1.173. Effective mAs [22] ranged from 86 to 206 (mean ± SD, 164.5 ± 19.3) using automatic tube current modulation (Dose-Right, Philips Medical Systems).
The raw projection data were reconstructed into 5-mm transverse sections at 4-mm intervals (thick-section). One image was randomly selected per patient to form a 100-image set. These images included 58 sections at or above the umbilicus and 42 sections below the umbilicus. At the same z-axis positions from the original raw projection data, 0.67-mm-thick (thin-section) images were reconstructed. All other reconstruction parameters were kept constant between each pair of thin- and thick-section images: image position along the x- and y-axes; field-of-view, 249-389 mm; matrix size, 512 x 512; and reconstruction filter type, filter B.
Image Compression
Each of the 100 thick- and 100 thin-section images had a bit depth of 12
bits per pixel packed into two bytes using a JPEG 2000 algorithm (Pegasus
Imaging), which is used in many commercial PACS. Each original image was
compressed to four different levels: reversible (as a negative control) and
irreversible 6:1, 10:1, and 15:1. These compressed images were then
decompressed, yielding 800 compressed (and then decompressed) images (thin- vs
thick-section x 100 images x four compression levels) for
comparison with the original images. The JPEG 2000 encoder was set to default
settings: 5-3 wavelet filter for reversible compression and 9-7 wavelet filter
for irreversible compression; single tile; six levels of wavelet
decomposition; size of code-block, 64 x 64; size of precinct, 32,768
x 32,768; 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 2.75 ± 0.20 (mean ± SD), 6.00
± 0.02, 10.00 ± 0.07, and 15.00 ± 0.11 for the
thin-section images and 3.64 ± 0.28 (mean ± SD), 6.24 ±
0.41, 10.00 ± 0.08, and 14.98 ± 0.11 for the thick-section
images. These variations from the nominal levels were considered unimportant
in this study.
PSNR
After conversion of the images to 8-bit images by applying a window setting
(window level, 20 H; window width, 400 H), PSNR (dB) was calculated for each
pair of original and irreversibly compressed images. The following equation
was used:
 |
 |
where RMSE stands for root-mean-square error and f(x, y) and g(x, y) are the pixel values in the original and compressed images, respectively.
Visual Analysis
Four board-certified body radiologists with 8, 7, 5, and 4 years of working
experience in interpreting body CT findings (reviewers 1-4) and a third-year
radiology resident with a 5-month rotation of abdominal imaging experience
(reviewer 5) participated.
Each of the 800 compressed images was paired with its original image for visual comparison. The 800 image pairs were randomly assigned to 16 reviewing sessions, avoiding repetition of any patient in a session. The order of reviewing sessions changed for each reviewer. Each session was separated by a minimum of 1 week.
Each image pair was displayed on a single monitor in an alternating fashion, and the order of the original and compressed images was randomized. The reviewer selectively toggled between the two images and could return to the first image as desired. Each reviewer independently determined if the two images were identical (indistinguishable) or if any detectable difference was present between them (distinguishable). When making comparisons, the reviewers were asked to pay attention to structural details, particularly small vessels and the edges 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 viewing software (PiViewSTAR version 5.0.7, SmartPACS), a flat-panel monochrome monitor (ME315, Totoku) with a matrix size of 1,536 x 2,048 and a diagonal display size of 20.8 in (52.8 cm) and matching video hardware (LV32P1, Totoku). The display system was calibrated to approximate the Barten tone scale [23] by using software (Medivisor Gray-Scale, Totoku) and a luminance meter (Minolta LS-110, Konica Minolta). The maximum and minimum luminances were 408.8 and 0.8 cd/m2, respectively. The ambient light of the room was subdued. All annotations and labels suggesting the compression level or scanning parameters were toggled off. Images were initially presented with a specific window setting (window level, 20 H; window width, 400 H), but the reviewers were allowed to adjust the window setting. Magnifying images was not allowed. Reviewing distance was constrained to a range of 34 to 78 cm by aiming a laser beam in front of each reviewer's forehead onto a ruler perpendicular to the monitor screen. The reviewing distance range of the reviewers had been measured during 30 minutes of their clinical work. Reviewing was conducted at the reviewers' convenience, without time constraint.
Statistical Analysis
A biostatistician participated in the study design using StatsDirect
statistical analysis software (StatsDirect). For each compression level, PSNR
was compared between the thin- and thick-section images by using paired
Student's t tests.
Interobserver agreements for the 400 thin-section image pairs and for the 400 thick-section image pairs were measured using kappa statistics for multiple reviewers [24]. The reviewers' responses were pooled: if three or more reviewers rated an image pair as distinguishable, then the pooled reviewers' response was considered to be distinguishable; otherwise, the response was considered indistinguishable. For each reviewer's response and for the pooled reviewers' responses, the percentage of distinguishable pairs was compared between thin- and thick-section images at each compression level using the exact test for paired proportions [25]. For the pooled reviewers' responses, a p value less than 0.05 was considered statistically significant. For each reviewer's response, the p value threshold was adjusted to 0.01 by using the Bonferroni method because five reviewers analyzed the same data set.
To estimate the visually lossless thresholds for thin- and thick-section images, we compared the percentage of distinguishable pairs from each of the three irreversible compression levels to the reversible compression (as a negative control). This analysis was performed for each reviewer's response and for the pooled reviewers' responses, by using the exact test for paired proportions. If the null hypothesis was rejected at a compression level, the visually lossless threshold was regarded as being below this compression level; otherwise, it was regarded as being above the compression level. Because this analysis involved three comparisons among the compression levels (each of the 6:1, 10:1, and 15:1 compressions compared with the reversible compression), the p value threshold was adjusted to 0.017 using the Bonferroni method.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Reviewers' responses are tabulated in Table 1. Kappa statistics were 0.79 and 0.73 for the thin- and thick-section images, respectively, indicating good interobserver agreement [26]. According to the pooled responses, at reversible compression, none of the thin- and thick-section images were distinguishable from the originals (p value could not be calculated ["null"] because the two compared percentages were both 0); at 6:1 compression, 27% (27/100) of the thin- and 0% of the thick-section images were distinguishable (p < 0.001); at 10:1 compression, 100% of the thin- and 80% of the thick-section images were distinguishable (p < 0.001); and at 15:1 compression, all of the thin- and thick-section images were distinguishable (p value could not be calculated because the two percentages were both 100) (Fig. 1A, 1B; for comparison, see supplementary Figs. S1C-S1J at www.ajronline.org).
|
|
|
Except for one reviewer (reviewer 5), the visually lossless threshold for the thin-section images was regarded as being below 6:1 because the percentages of distinguishable pairs at the 6:1 (12-63%), 10:1 (99-100%), and 15:1 (100%) compressions were significantly greater than that of the reversible compression (0-2%) (p < 0.001, for reviewers 1-4 and for pooled reviewers). For reviewer 5, the visually lossless threshold for thin-section images was regarded as being between 6:1 and 10:1 because the percentage at 6:1 (9%) was not significantly different from that of the reversible compression (1%) (p = 0.022, larger than the adjusted p value threshold of 0.017), whereas the percentages increased significantly at the 10:1 (98%) and 15:1 (98%) compressions (p < 0.001). The visually lossless threshold for thick-section images was regarded as being between 6:1 and 10:1 because the percentage at 6:1 (0-8%) was not significantly different than that of the reversible compression (0-4%) (p = 0.125-null for each reviewer; p = null for pooled reviewers), whereas the percentages increased significantly at the 10:1 (46-95%) and 15:1 (94-100%) compressions (p <0.001 for each reviewer and for pooled reviewers).
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
These results are not surprising because CT images with a thinner section are grainier because of more random noise caused by fewer photons contributing to the reconstruction of a voxel [27]. The noise is known to be vulnerable to compression because its energy is spread over numerous small coefficients in the wavelet domain and therefore is largely discarded even at low compression levels such as those used in this study [6, 19, 28].
Although we did not analyze the patterns of the artifacts but focused only on their magnitude, we have the impression that the perceived artifacts, which were more pronounced in thin-section images, have a blurring effect mainly altering the textures of the solid organs and soft tissues (Fig. 1A, 1B). This denoising effect at low compression levels might be negligible from a diagnostic viewpoint and may even be preferred from an esthetic viewpoint [6]. However, the effects of these artifacts on clinical interpretation are beyond the scope of this study. Many studies would be required to determine if the de facto compression level for thick-section images [7-15] is diagnostically acceptable for thin-section images. Until a consensus on compression level for thin-section images is reached to cover a broad range of potential abnormalities in the abdomen, we believe using a lower compression level would be more prudent for thin-section images.
In this study, the tested images had relatively homogeneous scanning and reconstruction parameters apart from section thickness. Provided that the image noise might be an important determinant of the compression tolerance of a CT image, more studies are needed to gain insight into the separate effects on the compression tolerance of various other imaging parameters affecting the image noise (body size, tube potential, tube current, use of automatic tube current modulation, and reconstruction kernel) [27].
Of the many CT parameters possibly affecting compression tolerance, we focused on section thickness in this study because it is one of the most advantageous features of modern CT scanners over previous scanners. Newer scanners can generate images of sub-millimeter thickness in routine abdominal applications. Radiologists now have the flexibility to change the section thickness after the acquisition is finished. Many radiologists are even routinely reconstructing two complementary data sets of thin and thick sections from a single acquisition [1, 29] to emphasize spatial resolution along the z-axis and low contrast resolution, respectively [30].
To our knowledge, there are only two previous studies in which the effect of section thickness on the compression of CT images was evaluated. Yamamoto et al. [31] reported that 2-mm-thick high-resolution lung CT images are more vulnerable to compression than 10-mm-thick conventional lung CT images on the basis of the comparison of PSNR and subjective rating of the image quality. In that study, a JPEG algorithm was used, the CT images had an 8-bit depth, only regions of interest (ROIs) (64 x 64 pixels) were compressed, and the study sample was small (28 ROIs from eight patients).
Siddiqui et al. [17] and Siegel et al. [18] recently reported that thinner sections are less compressible, based on their observations of the PSNR and a computer-based perceptual metric [32] in five chest CT data sets compressed with the JPEG 2000 algorithm. They also showed this reduced compressibility of thin sections can be compensated using the JPEG 2000 3D (part 2) algorithm, an extension of JPEG 2000 for 3D data, which takes advantage of correlation between adjacent images. To confirm these previous observations, more rigorous studies in the abdomen using formal visual analysis would be worth-while because both thin (for spatial resolution) and thick sections (for low contrast resolution) are important in abdominal CT [29, 30], and using the thinner (and therefore greater number of) images will foster the drive to adopt irreversible compression.
The visually lossless thresholds estimated in this study are likely to be smaller than the diagnostically lossless thresholds reported in previous studies on CT data compression [7-15]. Our study design was intended to be as conservative as possible in any estimate of the acceptable compression level by measuring visually lossless threshold with an image comparison method that is highly sensitive to image differences [20, 33-35]. During visual analysis, the reviewers might have learned characteristics of artifact patterns that are not clinically important and then exploited them in making decisions, which is a different process from real diagnostic interpretation. This seems an inevitable limitation common to any investigation on visually lossless threshold [16, 20, 21, 33]. Therefore, the visually lossless threshold measured in this study should be regarded as the minimum (baseline) acceptable compression level rather than an optimal compression level in clinical practice. Although the visually lossless threshold admits a relatively lower compression level, this conservative criterion has been gaining support as a practicable compression level [21, 33, 36] because it is more readily acceptable even by skeptical radiologists and the study results can be generalized more easily, regardless of diagnostic tasks [16, 20, 33].
This study has limitations. First, it was impossible to completely blind the reviewers to whether the tested image was a thin- or thick-section because they could frequently guess it from the graininess in the image. Therefore, the measured difference in perceptible artifacts between thin- and thick-section images might be exaggerated. 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 image comparison method was sensitive enough to detect minute perceptible compression artifacts regardless of the image content. Third, to avoid a possible clustering effect in a statistical sense, we tested only a single image per patient, which is unlike the clinical situation in which radiologists scroll through a series of images. A more elaborate study is necessary to approximate a real clinical situation. Fourth, the thin sections are not usually reviewed in clinical interpretations. Given that they are rather used as source data sets for 3D reconstructions [1], it would have been interesting to investigate the effects of compressing the source data set on the final 3D reconstructions. Nevertheless, the fidelity of the final 3D images would depend on the fidelity of the compressed source data set, which was focused on in this study.
In conclusion, compared with thick-section abdominal CT images, thin-section images are less tolerant to irreversible JPEG 2000 compression, and a lower compression level should be used for the visually lossless threshold.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
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]
|
|
|
|
 |
|
 H. Ringl, R. Schernthaner, E. Sala, K. El-Rabadi, M. Weber, W. Schima, C. J. Herold, and A. K. Dixon Lossy 3D JPEG2000 Compression of Abdominal CT Images in Patients with Acute Abdominal Complaints: Effect of Compression Ratio on Diagnostic Confidence and Accuracy Radiology, August 1, 2008; 248(2): 476 - 484. [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]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
