| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Technical Developments |
1 From the Departments of Radiology (L.M.F.) and Biomedical Engineering (Y.J., A.F.L.), Columbia University, New York, NY; and Department of Radiology, New York Presbyterian Hospital/Columbia University, New York, NY (L.M.F., Y.M.B., G.D.P., B.F., R.V.H.). Received May 21, 2001; revision requested July 9; final revision received December 10; accepted January 22, 2002. Address correspondence to L.M.F., Department of Radiology, Thomas Jefferson University, 132 S 10th St, Suite 1096, Philadelphia, PA 19107 (e-mail: lmf200@hotmail.com).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMaterials and Methods ResultsDiscussionREFERENCES |
|---|
© RSNA, 2002
Index terms: Computed tomography (CT), image display and recording, **.121192, 60.12119, 76.12119 • Computed tomography (CT), image processing, **.12119, 60.12119, 76.12119 • Liver, CT, 76.12119 • Lung, CT, 60.12119
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionREFERENCES |
|---|
The current state-of-the-art method for automatic display is a variation of the technique of histogram equalization (3–6). Histogram equalization is a nonlinear transformation scheme that maps image intensity values across the entire range of a display device. In 1985, Lehr and Capek (6) studied histogram equalization enhancement of CT images and showed that although there was no decrease in detectability of simulated liver metastases, radiologists found that histogram equalization markedly enhanced the display of image noise and artifacts. Subsequently, Pizer et al (5) described a method called adaptive histogram equalization, or AHE, that applies the concept of histogram equalization to small overlapping local areas of the image. This method produced "attractive results, as judged by many radiologists" (5). To our knowledge, however, no study was published that described the diagnostic capability of adaptive histogram equalization to support such enhancement of CT images as an automatic display.
Contrast-limited adaptive histogram equalization (CLAHE) is a refinement of adaptive histogram equalization (3,5,7). In CLAHE, unlike in adaptive histogram equalization, the amount of contrast enhancement that can be produced within a local area of the image is restricted by an adjustable parameter known as the "clipping level." In 1989, Zimmerman et al (3) showed that physicians rated CLAHE as providing superior subjective image quality compared with adaptive histogram equalization, but they found no difference between CLAHE and adaptive histogram equalization in their ability to depict subtle lung nodules on an image. In the latter study, however, images enhanced with adaptive histogram equalization were compared with CLAHE-enhanced images only. No comparison was made with conventional display.
Unlike histogram equalization techniques, wavelet-based enhancement algorithms offer the opportunity for multiscale or multiresolution analysis. With wavelet theory, an image can be decomposed into a series of images at different levels of spatial resolution. In this way, an image can be manipulated and enhanced at each of its fundamental resolution levels. Subsequent reconstruction and reassembly of the decomposed levels may allow the enhancement of a particular feature on an image. As a result, wavelet algorithms offer the ability to display differing degrees of contrast and detail on an image and allow the full dynamic range to be displayed simultaneously in one automatic window.
A wavelet-based CT display algorithm, multiscale adaptive histogram equalization (MAHE), has recently been described as an automatic method of CT image display (8,9). The purpose of this study was to determine whether individual MAHE-enhanced CT images can be interpreted faster than can those displayed with conventional window settings, without compromising the detectability of radiologic abnormalities. Our study had two objectives: (a) to compare MAHE image interpretation times with those of conventional window display and (b) to evaluate the diagnostic performance of MAHE.
|
Materials and Methods |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionREFERENCES |
|---|
Test Algorithms
CLAHE.—Our team decided the optimal parameters for implementation of CLAHE after an extensive evaluation of reported advanced histogram equalization algorithms, including those described for the evaluation of mammograms, skeletal lesions, nuclear medicine studies, and chest radiographs (10–15). The CLAHE algorithm we used was supplied by Zuiderveld (16), with use of a contextual region size of 32 pixels per bin and a clip level of 3 pixels per bin (3,9). Figure 1a and 1b show images displayed with conventional window settings and after CLAHE enhancement, respectively.
|
|
|
|
Patients underwent helical CT (Somatom 4 Plus; Siemens Medical Systems, Iselin, NJ). Scanning parameters were 120 kV, 280 mA, and 5-mm-thick sections. Of 109 images, 104 were obtained without intravenous injection of contrast material. The images, each 512 x 512 pixels, were transferred over an electronic network to a workstation (IntelliStation Z-Pro with dual PII, 430-MHz Xeon; IBM, Armonk, NY), where they were saved as raw files without prior preprocessing, bypassing the ancillary header information included in the original digital imaging and communications in medicine, or DICOM, format.
Each raw file was subsequently processed with the two test algorithms. Hence, three complete sets of the database images were obtained: the original images, the CLAHE-enhanced images, and the MAHE-enhanced images. These test images were transferred by means of the Ethernet to a workstation (Ultra Sparc, Solaris, version 8; Sun Microsystems, Mountain View, Calif), where they were saved for viewing by the readers. The workstation was in the main reading room in the department of radiology and had optimal surrounding ambient light conditions. The images were presented on a 20-inch (51-cm) high-resolution monitor by using software (Medx, version 2.0; Sensor Systems, Sterling, Va), with a 20 x 20-cm field of view.
Readers and Procedures
Three board-certified radiologists, one general (B.F.) and two specialized chest (Y.M.B., G.D.P.) radiologists, participated in this study. The readers had the following levels of experience: reader 1, 30 years; reader 2, less than 10 years; and reader 3, 1 year. Hence, the effect of experience with CT image interpretation could be examined. The readers worked separately and interpreted the study images independently, without a consensus interpretation.
The readers were shown the CLAHE- and MAHE-enhanced images and the unprocessed original images as the control standard. In this way, each observer could serve as his or her own control. The three sets of images were shown to each reader in three sessions separated by at least 2-week intervals, to avoid recall bias (Huda W, oral communication, 2000). In each session, the reader viewed images of only one type: processed or original. Images displayed with conventional window settings were shown first, followed by CLAHE-enhanced images, and then MAHE-enhanced images. The images at each session were shown in random order, to avoid biases of reader order effects.
The images were shown sequentially to the radiologists by one observer (L.M.F.), who recorded the time required for the interpretation of each image. In the case of the control situation, interpretation time was recorded for each image, which was shown with three conventional window center and level settings: 29 and 350, -450 and 1,500, and 300 and 1,500 for mediastinal, lung, and bone tissues, respectively.
During interpretation, the readers were allowed to vary their position and viewing distance as needed. Each reader was allowed as much time as he or she required to fully interpret each CT image.
Each reader was given a training session to accustom the reader to the specifics of each algorithm and to orient the reader to the study procedures. A set of training images was shown that contained representative abnormalities in the lung, mediastinum, and bones displayed with conventional window settings, MAHE enhancement, and CLAHE enhancement. Images used for training were not included in the set of 109 study images. Each reader was allowed as much time as he or she required to examine the training images.
Each reader was informed that each image might contain no lesions, one lesion, or multiple lesions in the lungs, soft tissues, and bones. The readers were first asked to rate the quality of each image on a scale from 1 to 5: 1, very poor; 2, poor; 3, weakly satisfactory; 4, good; and 5, excellent. Next, they rated the likelihood that a lesion was present on a scale from 1 to 3: 1, definitely present; 2, likely present; and 3, equivocally present. Confidence limits were introduced into this study for the purpose of quantifying the readers’ confidence with respect to the detection of abnormalities on the images processed with the test algorithms. Since each reader served as his or her own control, the readers used their own internal standards for confidence level characterization.
The readers were asked to characterize their observations as follows: Lung abnormalities included air cysts, nodules with a diameter of 1 cm or less, nodules larger than 1 cm, ground-glass opacities, consolidation, and linear opacities. Mediastinal abnormalities were defined as lymph nodes smaller than 1 cm, lymphadenopathy (lymph nodes larger than 1 cm), or pericardial thickening or effusion. Pleural disease included thickening and effusion. Bone lesions were defined as sclerotic, lytic, or mixed lesions. Finally, the lesion location was recorded to ensure interreader correlation of observations. The readers were instructed to indicate a lesion’s location in the lung on the basis of the quadrant: quadrant 1, right anterior; quadrant 2, left anterior; quadrant 3, right posterior; and quadrant 4, left posterior. Readers indicated a lesion’s location in the mediastinum on the basis of the location of the lymph node station or pericardial involvement. In the bones, the readers indicated a lesion in one of four locations: right ribs, left ribs, vertebra, or sternum.
Analysis
To compare the time required to interpret each image displayed with conventional window settings, CLAHE enhancement, and MAHE enhancement, a statistical repeated-measures analysis of variance was used (18).
To compare the overall diagnostic capability of each test algorithm, a consensus of two of the three readers’ interpretations of each of the algorithm-enhanced set of images was used with confidence levels of 1, definitely present, and 2, likely present, and compared with the consensus interpretation obtained with conventional window settings. Histologic correlation was not provided in this study; therefore, overall diagnostic truth was determined with a consensus of two of the three radiologists’ interpretations of each image viewed with the conventional window settings at confidence levels of 1 and 2. Hence, diagnostic sensitivity for conventional window settings was 100%.
Detection rates of individual readers were also compared to discover possible diagnostic advantages for the display algorithms over conventional window settings. The latter comparison was performed despite the absence of histologic correlation, which necessitated that any apparent improvement in diagnostic sensitivity for a test algorithm could be suggested but not definitively made. In addition, individual reader detection rates were calculated for all confidence levels and compared to uncover the readers’ confidence with respect to interpretation of the processed images.
Statistical significance of the consensus readings for each algorithm-enhanced set of images compared with the images displayed with conventional window settings was determined with a 
2 test (19–21). A McNemar test was then used to determine the significance of the difference between the three display modes for the consensus readings and the individual readers’ observations (22).
Diagnostic sensitivity and specificity were calculated for all pulmonary, pleural, mediastinal, and bone lesions. For pulmonary lesions, diagnostic sensitivity and specificity were calculated for specifically characterized lesions (air cysts, nodules, ground-glass opacities, linear densities, and consolidation) and lung lesions, regardless of characterization, to account for interreader differences in characterization. In addition, detection rates for small and large nodules were compared among the three display algorithms.
Differences with a P < .05 were considered statistically significant.
|
Results |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionREFERENCES |
|---|
|
For images displayed with conventional window settings, the quality of the images was rated as follows: reader 1, 101 of 109 (92.7%) excellent, eight of 109 (7.3%) good; reader 2, 31 of 109 (28.4%) excellent, 78 of 109 (71.6%) good; reader 3, 100% excellent. For all images displayed with each of the test algorithms, the readers uniformly had difficulty rating the quality of the images and offered no quality ratings for these images.
Tables 3 and 4 show overall diagnostic sensitivities and specificities obtained with the consensus of two of the three readers’ interpretations of the three sets of images. Table 5 shows overall diagnostic sensitivities and specificities obtained with consensus of two of the three readers’ interpretations of small (
1 cm) and large (>1 cm) nodules.
|
|
|
|
|
|
|
|
|
|
No statistically significant difference was observed among the three confidence levels for any reader regarding all lesions seen with the three display modes.
Informal reader observations include that MAHE-enhanced CT images of the chest show fine interstitial markings in the periphery of the lung, where conventional window settings show no vasculature.
|
Discussion |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionREFERENCES |
|---|
Initial experimentation with histogram equalization techniques has shown limited success. Particularly, a recognized weakness of CLAHE enhancement is that small regions of sharp contrast change on an image are deemphasized (3–5). As a result, in the lung, all structures (pulmonary vessels and nodular opacities) displayed in the background of air in the alveoli are enlarged and slightly distorted. The latter principle may explain why, for one reader, there was an apparent increased detection rate of small nodules on CLAHE-enhanced images compared with that on images displayed with conventional window settings. On the other hand, a region of ground-glass opacity will lack detail and appear as an indistinct relatively homogeneous opacity. Thus, for the detection of nonspecific opacification in the lungs, CLAHE is useful, but it is limited for detailing the characteristics of a lung opacity.
In the mediastinum, structures surrounded by fat are well outlined, in concert with the principle that opacities in the lung are enlarged. Hence, sensitivity for the detection of mediastinal disease was comparable to that with conventional window settings for two of the three readers.
To our knowledge, this study is the first investigation that showed the diagnostic capability of CLAHE-enhanced CT images for a wide variety of pathologic conditions in the chest. Although no definite advantage for CLAHE enhancement over conventional window settings could be identified in this study, an open question remains as to its usefulness as an aid for the detection of small nodules.
MAHE was developed to circumvent and reduce some of the artifacts visualized with CLAHE enhancement. MAHE is a wavelet-based algorithm in which the wavelet expansions contain built-in redundancy of representation. Hence, this algorithm is less susceptible to artifacts and perturbations, unlike the exact representation of histogram equalization techniques, where a perturbation can be magnified and cause a catastrophic artifact.
MAHE-enhanced CT images of the chest show fine interstitial markings in the periphery of the lung, where images displayed with conventional window settings show no vasculature. We hypothesize that these intersititial markings are true findings brought out with the algorithm rather than noise, as contrast improvement is gained on an image without sacrificing detail. At this time, however, we have no histologic evidence to support this hypothesis.
MAHE is primarily a contrast-enhancing algorithm that accentuates all naturally occurring areas of contrast on the image. Therefore, contrast enhancement is best accomplished in the lung. With respect to specific abnormalities, the MAHE algorithm was superior to the CLAHE algorithm for the detection of air cysts (sensitivities of 85% compared with 52%, respectively). In addition, a statistically significant increase in the detection rate of air cysts with MAHE-enhanced images compared with images displayed with conventional window settings was seen for at least two of the three readers. This finding led to the suggestion that, in the absence of histologic correlation, MAHE may be useful clinically for the detection of air cysts. In the mediastinum and bones, where there are more homogeneous densities with fewer intervening attenuation changes, contrast enhancement is visually reduced compared with that obtained with the CLAHE algorithm. However, there was no statistically significant difference in the detectability of bone and mediastinal lesions.
For the detection of small nodules (1 cm) on MAHE-enhanced images, there was a decrease in detection rate compared with that observed with both CLAHE and conventional window settings for two of the three readers. For MAHE-enhanced images, unlike CLAHE-enhanced images, there is no distortion of size or configuration in the appearance of lung opacities. However, the enhanced interstitial markings provide a background against which nodules may be less easily visualized, which may explain the decreased detectability of small nodules. Likely for a similar reason, the detection of ground-glass opacities was significantly reduced with MAHE-enhanced images compared with images displayed with conventional window settings.
Our hypothesis that one automatic display window will require a shorter CT interpretation time was confirmed in this study. Essentially, the mean time required to interpret an MAHE-enhanced image (8.8 seconds) was less than that required to interpret the same image displayed with two conventional window displays. For the reading of an entire CT scan of the chest, the time savings for a CT study that contains 60 images will be approximately 3.3 minutes.
The definition of truth in this investigation was the consensus of two of the three radiologists’ interpretations of the CT images displayed with conventional window settings. We used this arbitrary definition because histologic correlation was not performed in this pilot study. Not surprisingly, overall diagnostic accuracy for both test algorithms, CLAHE and MAHE, is significantly lower than that with conventional window settings. However, analysis of the data from individual readers showed few statistically significant differences in performance between the test algorithms and conventional window settings, primarily for mediastinal, pleural, and bone lesions. Furthermore, for air cysts, there was an apparent increased detection rate compared with that with images displayed with conventional window settings.
Nevertheless, there are several biases that potentially compromised the results of this investigation. An important bias in this study was the overwhelming familiarity and experience that the readers had with the conventional window display and their equally overwhelming inexperience with the new test algorithms and the potential artifacts associated with them. Although the training sessions were designed to circumvent this bias, the lack of practice of each reader with the interpretation of images enhanced with the test algorithms is likely impossible to overcome.
An attempt was made in this study to control for the possibility of recall bias: In each session, study images were shown in random order. Also, although only one type of image was shown in each session (either processed or original), study sessions took place at 2-week intervals, to minimize recall bias. Nevertheless, a superior study design would have included all three types of image displays—CLAHE, MAHE, and conventional window settings— in the same session and shown in random order.
In part because of lack of experience with the test algorithms, the readers uniformly had difficulty rating the quality of the images displayed with each of them. For images displayed with conventional window settings, however, quality was rated between good and excellent among all the readers.
Interreader observations were complicated by several factors. Notably, because individual CT images were shown to each reader without the benefit of contiguous CT sections or a complete CT study, each reader may have characterized an opacity in a different manner. On several images, for example, a ground-glass opacity characterized by readers 1 and 2 was characterized as a nodular opacity by reader 3. This phenomenon is especially true in cases in which the opacity was complex and did not fit neatly into one category. To address this problem, an analysis of detection rates was also made on the basis of the total number of lung opacities visualized on each image, regardless of characterization. The analysis showed no suggestion of an advantage for display with either of the test algorithms over that with conventional window settings.
Although the readers were allowed as much time as they required to interpret each image and were allowed as many breaks as they needed to properly focus on the reading of each image, the study may have been compromised by potential boredom and fatigue of the reader. A superior study design may have included repeat study of each reader’s interpretation of the test images.
This pilot study tested the potential usefulness of CLAHE and MAHE in a clinical setting. Although the relative relationship of radiographic opacity on an image is preserved after enhancement with both test algorithms, one criticism of these algorithms is that attenuation values of structures are altered from their original values during the enhancement process. Fortunately, this problem can be overcome. Original attenuation values can be made available automatically by means of software correlation of the algorithm-enhanced CT image with its original image.
In conclusion, MAHE is a wavelet-based technique, inspired from adaptive histogram equalization, that provides a method for automatic simultaneous display of the full dynamic range of a CT image. Although we observed a significant reduction in interpretation time required for MAHE-enhanced images compared with that with conventional window settings, overall accuracy was insufficient in our pilot study to suggest MAHE as a potential replacement for conventional window settings. Investigation of lung abnormalities yielded mixed results, with a particularly good detection rate for air cysts, which leads to speculation that MAHE may be a useful tool for the detection of emphysema. Conversely, the detection rate for small nodules was poor. Unlike traditional histogram equalization techniques, however, MAHE is built on a platform of multiresolution analysis with use of redundant representations. This algorithm may be further manipulated and improved by using varied basis functions and by incorporating clipping levels into the adaptive equalization process, to specifically target challenging diagnostic problems, such as the detection of small nodules.
|
ACKNOWLEDGMENTS |
|---|
|
FOOTNOTES |
|---|
Abbreviations: CLAHE = contrast-limited adaptive histogram equalization, MAHE = multiscale adaptive histogram equalization
Author contributions: Guarantors of integrity of entire study, L.M.F., A.F.L., R.V.H.; study concepts and design, all authors; literature research, L.M.F., Y.J.; clinical and experimental studies, G.D.P., Y.M.B., B.F., L.M.F., Y.J., A.F.L.; data acquisition, L.M.F., Y.M.B., Y.J., G.D.P., B.F.; data analysis/interpretation, L.M.F., Y.M.B., Y.J., G.D.P.; statistical analysis, L.M.F., Y.J.; manuscript preparation, L.M.F., Y.J., A.F.L., R.V.H.; manuscript definition of intellectual content, L.M.F., Y.J., A.F.L., R.V.H., Y.M.B.; manuscript editing, L.M.F., Y.J., A.F.L., R.V.H., Y.M.B., G.D.P.; manuscript revision/review and final version approval, all authors.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionREFERENCES |
|---|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
