| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Original Research |
1 All authors: Department of Radiology, Seoul National University Bundang Hospital, 300 Gumi-dong, Bundang-gu, Seongnam-si, Gyeonggi-do, 463-707, Korea; and Seoul National University College of Medicine, Institute of Radiation Medicine, Seoul National University Medical Research, Seoul, Korea.
Received October 26, 2007;
accepted after revision February 23, 2008.
Supported by the Seoul R&BD Program, Republic of Korea (project number
10675).
Abstract
 Top Abstract IntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
MATERIALS AND METHODS. Thirty-five thin-section (1 mm) and 35 corresponding thick-section (5 mm) images were compressed to reversible and irreversible 4:1, 6:1, 8:1, 10:1, and 15:1. In each compressed image, pixels outside the lung were replaced with those from the original image. By comparing the compressed and original images, three radiologists independently rated the compression artifacts using grades of 0 (none, the two images were indistinguishable), 1 (image differences were barely perceptible), 2 (image differences were subtle), or 3 (image differences were significant). At each compression level, thin and thick sections were compared for peak signal-to-noise ratio (PSNR) using paired t tests and for the readers' responses using Wilcoxon's signed rank tests and exact tests for paired proportions.
RESULTS. Thin sections had smaller PSNR (p < 0.0001). Thin sections had higher grades of artifacts than thick sections, showing significant differences at compression levels of 10:1 (mean score, 0.8 vs 0.4, 0.9 vs 0.1, 0.3 vs 0.0; p < 0.009 for the three readers) and 15:1 (1.9 vs 1.0, 1.9 vs 1.1, 1.5 vs 0.6; p < 0.0001). The percentages of distinguishable pairs (grades 1–3) were greater for thin sections than for thick sections, showing a statistically significant difference at 10:1 for two readers (31% vs 3% and 74% vs 37%; p < 0.006).
CONCLUSION. The lung shows more compression artifacts on thin sections than on thick sections. Section thickness should be taken into consideration when adjusting the compression level for lung CT images.
Keywords: artifacts • CT • data compression • lung
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
The purpose of this study is to show the difference in the amount of Joint Photographic Experts Group (JPEG) 2000 compression artifacts in the lung between thin- and thick-section CT images with a lung window setting. We focused on the lung, which is the region of principal interest in images with the lung window setting.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Study Sample
A chest radiologist with 5 years of clinical experience compiled 35 lung CT
images obtained using a 64-MDCT scanner (Brilliance, Philips Healthcare)
during his clinical work in September 2006, according to the following image
selection scheme. One image was selected per patient to form a 35-image set,
including five normal and 30 abnormal images. In the 30 abnormal images, the
following six pathologic findings were distributed equally on five images
each: ground-glass opacity, interlobular septal thickening, air trapping,
micronodule, mass, and consolidation. The definitions and descriptions of
these findings were in accordance with the recommendations of the Fleischner
Society [18]. For each of
these six pathologic findings, each group of five images included two images
showing extensive disease, one showing moderate disease, and two showing
subtle disease. The selected patients were 22 men and 13 women ranging from 20
to 82 years in age (mean, 51 years). The selected images included 10 images
above the carina and 25 images below the carina. The images were from 18
contrast-enhanced and 17 unenhanced scanning sequences.
CT Scanning
The scanning parameters were detector collimation, 0.625 mm; gantry
rotation time, 0.5 second; tube potential, 120 kVp; and pitch,
1.078–1.173. Effective mAs ranged between 165 and 210 mAs (mean, 195.0
± 16.9 [SD] mAs) using automatic tube current modulation (Dose-Right,
Philips Healthcare). For routine clinical practice, the raw projection data
were reconstructed into 5-mm transverse sections at 4-mm intervals, from which
the aforementioned 35 images (thick sections) were selected. To match each of
these images, a 1.0-mm-thick image (thin section) was reconstructed at the
same z-axis position from the raw projection data. 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 (278–390 mm), matrix size (512 x
512), and reconstruction filter type (filter C).
Image Compression
Each of the 35 thick and 35 thin sections had a bit depth of 12 bits per
pixel packed into 2 bytes. Compression level was defined as the ratio of the
original pixel size (16 bits/pixel) to compressed size in bits per pixel
[19]. Using a JPEG 2000
algorithm (PICTools, version 2.00.543, Pegasus Imaging) that is used in many
commercial PACS, each image was compressed reversibly and irreversibly with
ratios of 4:1, 6:1, 8:1, 10:1, and 15:1. This yielded 420 (thin and thick
x 35 images x six compression levels) compressed (and then
decompressed) images for comparison with their originals. The JPEG 2000
encoder parameters were set to default settings: 5–3 (for the reversible
compressions) or 9–7 (for the irreversible compressions) wavelet filter,
single tile, six levels of wavelet decomposition, 64 x 64 code-block,
32,768 x 32,768 precinct, and a single layer. Minute differences of
actual compression levels from the nominal compression levels were considered
unimportant. The actual compression levels achieved with the reversible
compressions were 2.7 ± 0.1 (mean ± SD) and 3.2 ± 0.1 for
the thin and thick sections, respectively.
Masking the Mediastinum, Chest Wall, and Background Air
Because this study focused on the lung—the main region of interest in
images at the lung window setting—the chest wall, mediastinum, and
background air were masked to avoid their potential influences on subsequent
analyses. This was important because for a given compressed image, more
artifacts are usually discernable in the chest wall and mediastinum than in
the lung [6,
20]. On each original thick
section, a chest radiologist marked the lung silhouette by carefully drawing
lines just outside the pleura using a graphic tablet (Graphire pen tablet,
Wacom Technology), consequently leaving the lungs untouched. The lung
silhouette was then superimposed over the corresponding thin- and
thick-section compressed images. In these compressed images, the pixels in the
mediastinum, chest wall, and background air outside the silhouettes were
replaced with the corresponding pixels from the original image. Therefore, in
each resulting image, the pixels inside the lungs were from the compressed
image, and the pixels outside the lungs were from the original image. A bit
depth of 12 was maintained throughout this procedure. Hereinafter, these
masked compressed images are referred to as "compressed
images."
PSNR
After converting the original and compressed images to 8-bit images by
applying a lung window setting (level, –600 H; width, 1,500 H), PSNR
(dB) inside the lung region was calculated for each pair of original and
irreversibly compressed images. The following equation was used:
 |
 |
Image Analysis
Three board-certified radiologists with 8, 7, and 5 years of clinical
experience in reading thin-section chest CT scans participated in the image
analysis. Each of the 420 compressed images was paired with its original image
for visual comparison. The 420 image pairs were randomly assigned to 14
reading sessions while avoiding the repetition of any patient in a session.
The order of reading sessions changed for each reviewer. Each reading session
was separated by a minimum period of 3 weeks.
|
Images were displayed in a one-by-one format using viewing software (Pi-view Star version 5.0.7, Smart PACS), a flat-panel monochrome monitor (ME315, Totoku) with a matrix size of 1,536 x 2,048 and a diagonal display size of 52.8 cm, and matching video hardware (LV32P1, Totoku). The display system was calibrated to approximate the Barten tone scale [21]. The maximum and minimum luminance measured 393.3 and 1.2 cd/m2, respectively, with the ambient light of the room subdued. To reproduce clinical practice, the reading distance was constrained to a range of 32–78 cm by aiming a laser beam in front of each reader's forehead onto a ruler perpendicular to the monitor screen. The reading distance of the readers had been measured during 30 minutes of their clinical work. Otherwise, image review was conducted at each reader's convenience, without time constraints.
|
|
|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
The readers' responses are illustrated in Figures 2A, 2B, and 2C. Kappa statistics were 0.43 and 0.32 for the thin and thick sections, respectively, indicating fair interobserver agreements [24]. As the compression level gradually increased, the readers rated more compressed images as having higher grades of artifacts. Although two of the three readers assigned a grade of 3 to 23% (8/35) and 6% (2/35) of the 15:1 compressed thin sections, none of the compressed thick sections were rated as grade 3 at any compression level. At each compression level, the readers tended to assign higher grades of artifacts to the thin sections than to their corresponding thick sections (Fig. 3 and Fig. S3 in supplemental data at www.ajronline.org). As the compression level gradually increased, this difference became more apparent, showing statistical significance at compression levels of 10:1 (mean score, 0.8 ± 0.5 vs 0.4 ± 0.5, p = 0.004; 0.9 ± 0.7 vs 0.1 ± 0.3, p < 0.0001; and 0.3 ± 0.5 vs 0.0 ± 0.2, p = 0.009, for thin vs thick by the three readers, respectively) and 15:1 (1.9 ± 0.9 vs 1.0 ± 0.6, 1.9 ± 0.5 vs 1.1 ± 0.7, and 1.5 ± 0.6 vs 0.6 ± 0.6; p < 0.0001 for thin vs thick by all three readers). Likewise, the percentage of distinguishable pairs (grades 1–3) also tended to be greater for the thin sections than for the thick sections at each compression level. This difference became apparent at 10:1: The difference was statistically significant for reader 2 (69% [24/35] vs 9% [3/35], p < 0.0001) and reader 3 (31% [11/35] vs 3% [1/35], p < 0.006) and almost reached statistical significance for reader 1 (74% [26/35] vs 37% [13/35], p = 0.05). At 15:1, the statistical significance disappeared as all readers rated most of the image pairs as distinguishable for both thin and thick sections.
|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Although these results corroborate those of previous studies [4, 25–27] that mentioned that thin-section CT images are more vulnerable to compression than thick sections, several major differences in this study should be noted. The differences include the image selection scheme [4, 25–27], compression algorithm [27], compression levels [25–27], use of radiologists' artifact grading [4, 25, 26], viewing conditions [25–27], and anatomic regions [4]. These experimental settings in our study were chosen to match current lung CT practice. To our knowledge, our study is the first attempt to show that radiologists perceive different compression artifacts in thin- and thick-section lung CT images by using formal visual analyses.
In this study, the analyses were limited to the lung region by masking the chest wall, mediastinum, and background air. This focused attention on the lung, the region of principal interest in images at lung window settings. It has been recently revealed that radiologists perceive more artifacts in the chest wall and mediastinum than in the lung [20]. Therefore, if the region outside the lung had not been masked, the overall compression artifacts of a tested image might have been rated according to the artifacts in the chest wall and mediastinum, rather than the artifacts in the lungs. To avoid this, Ringl et al. [6] used rectangular collimation in their analysis of compression artifacts in the lung. However, in our study, the region outside the lung was replaced with pixels from the original image. This simulated all viewing conditions (including luminance adaptation in radiologists' visual system) in clinical interpretation.
Our results are as expected; thinner sections have less correlation between adjacent pixel values to be exploited by image compression, due to more random noise [28]. As the compression level gradually increases, the disappearance of the noise is known to be one of the first perceivable changes in an image [10]. Therefore, the artifacts rated as grade 1 or 2 in this study probably correspond, in part, to the de-noising. The denoising effect at low compression levels might be negligible from a diagnostic viewpoint and might even be preferred from an esthetic viewpoint [11], especially for noisier thin sections. However, note that, in the same compressed image, the de-noising effect is inevitably accompanied by blurring artifacts to some degree, altering the inherent pulmonary texture. To determine whether these minute artifacts can hinder diagnosis, larger studies using receiver operating characteristic analysis would be required. However, it seems unrealistic to conduct such studies to cover a broad range of pulmonary abnormalities in images with different section thicknesses.
Our study relied on the readers' subjective decisions in grading compression artifacts, as did studies by other researchers [6, 27, 29, 30]. The readers frequently assigned different grades to the same image. To be more objective and conservative, an additional analysis of the presence of any perceivable artifacts was conducted by collapsing the readers' grading into a binary variable. We believe this analysis was less subjective, although individual differences still existed in the readers' sensitivities in artifact perception. If a radiologist cannot distinguish a compressed image from its original, there is no basis for arguing that the compression hinders any diagnostic accuracy, regardless of image content. This criterion, called the "visually lossless" criterion, has been rapidly gaining support as a conservative and practical guideline for medical image compression [4, 6, 11, 31–34]. Despite the aforementioned individual variations and subjectivity in the visual analysis, the study results remain valid with reference to the artifact differences between thin and thick sections regarding their presence and relative magnitude.
In this study, we do not propose acceptable thresholds of compression level for thin- and thick-section lung CT images. The thresholds have been traditionally measured as the point at which the statistical significance disappears for the difference in diagnostic performance [5, 7, 29, 35, 36] or image quality [4, 6, 11, 20, 30, 32–34, 37, 38] between compressed and original images. Such an analysis is prone to statistical type 2 error [39], especially with the small sample size (35 tested images) used in this study. Even with a much greater sample size, determining a fixed compression guideline is still difficult because compression artifacts depend on many independent factors, including the compression algorithm, image acquisition parameters, and even image content [10, 11]. Therefore, the compression level should ideally be adjusted to individual images in an adaptive manner, probably using image quality metrics [32, 38]. Further investigations are warranted to develop such intelligent compression techniques and to validate their usefulness in medical image compression.
This study has other limitations. First, it was impossible to completely blind the readers to whether the tested image was a thin or thick section because they could frequently guess it from the graininess of the image. Therefore, the measured difference in perceptible artifacts between thin and thick sections might be exaggerated. Second, aside from the section thickness, the images we tested were of relatively homogeneous scanning and reconstruction parameters. Assuming that image noise may be an important determinant in the compression artifacts of a CT image [4], more studies are needed to gain insight into how the compression artifacts are separately affected by various other imaging parameters (body size, tube potential, tube current, use of automatic tube current modulation, and reconstruction kernel) affecting the image noise [28]. Third, because a limited number of images were tested, we could not analyze whether different pathologies were differentially affected by the compression. Although a larger study is needed for this purpose, note that the sample size in this study was large enough to prove the artifact differences between thin and thick sections.
In conclusion, the lung shows more JPEG 2000 compression artifacts on thin-section CT images at a lung window setting than on thick-section images. Section thickness should be taken into consideration when adjusting the clinical compression level for lung CT images.
Acknowledgments
We thank the radiologists who participated in our experiment.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
) sections.
