| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Computer Applications |
1 From the Departments of Medicine (C.K., D.S.P.) and Radiology (C.F.B., R.B.J., S.N.), Stanford University School of Medicine, Lucas MRS Center P-268, Stanford, CA 94305-5488. Received July 20, 1998; revision requested September 24; revision received November 18; accepted January 11, 1999. Supported in part by National Institutes of Health grants 1R01 CA72023, 1P41 RR09784-01, and LM 07033, the Packard Foundation (Los Altos, Calif), the Lucas Foundation (Menlo Park, Calif), and the Phil N. Allen Trust (Menlo Park, Calif). C.F.B. is a 1997 RSNA Scholar. C.K. is a 1998 GENDEX/RSNA Medical Student/Scholar Assistant. Address reprint requests to S.N. (e-mail: snapel@stanford.edu).
 |
Abstract |
|---|
 TOP Abstract IntroductionMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
MATERIALS AND METHODS: Spiral CT simulator software was used to generate 10 synthetic polyps in various configurations. Additional software was developed to insert these polyps into volume CT scans. Ten polyps in eight patients were selected for comparison. Three radiologists evaluated whether two-dimensional (2D) CT images and three-dimensional (3D) volume-rendered CT images showed synthetic or real polyps.
RESULTS: Edge-response profiles and noise of simulated polyps matched those of native polyps. Frequency distributions of reviewers' responses were not significantly different for synthetic versus real polyps in either 3D or 2D images. Responses were clustered around the response of "unsure" if lesions were real or synthetic. Receiver operating characteristic curves had areas of 0.54 (95% CI = 0.39, 0.68) for 3D and 0.39 (95% CI = 0.25, 0.53) for 2D images, which were not significantly different from random guessing (P = .70 and .28 for 3D and 2D images, respectively).
CONCLUSION: Synthetic polyps were indistinguishable from real polyps. This method can be used to generate ground truth experimental data for comparison of CT colonographic display and detection methods.
Index terms: Computed tomography (CT), image display and recording, 75.12115, 75.12117 • Computed tomography (CT), three-dimensional, 75.12117 • Computers, simulation • Colon, CT, 75.12111, 75.12115, 75.12117 • Colon, neoplasms, 75.3119
|
Introduction |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
Thus far, most investigators (2–5) have used a combination of two-dimensional (2D) and three-dimensional (3D) display modes. When 2D and 3D modes are used together in an unblinded fashion, it is difficult or impossible to establish accurately the relative diagnostic contribution of each mode. The authors of one initial study (2) suggested that 3D display added incremental value over 2D display alone, by increasing the sensitivity for lesion detection. However, more recent data (6,7) suggest that 2D display of "source" axial CT images alone may be sufficiently sensitive for lesion detection.
Several 3D display methods have been used (8–12), and more are under development. These approaches differ from one another not only in terms of the way the images are presented but also in their relative computational complexity and efficiency of viewing. Thus, an important remaining issue in the development of CT colonography for application to screening populations is to determine an efficient yet accurate means of reviewing the acquired image data for the detection of colonic polyps.
To optimize these display tools, a large number of data sets with known true polyps are needed. Typically, one compares CT colonographic images with fiberoptic colonoscopic images, the current reference standard, in the same patients. There are a number of difficulties with this approach. In screening populations, few patients have lesions in the clinically relevant size range of larger than 1 cm in diameter (1). In addition, fiberoptic colonoscopy is less than perfect, in part because the instrument fails to reach the cecum 5%–10% of the time (13). The colon also "telescopes" during the procedure, which makes precise reconciliation of CT findings with fiberoptic colonoscopic findings difficult. Last, because only a small number of lesions are available, radiologists may be able to recall specific lesions seen with different display modes according to shape or local surroundings and, hence, bias results.
A reference standard better than that available with fiberoptic colonoscopy and a large collection of polyps are needed for rigorous evaluation and optimization of CT colonography. The purpose of our study was to develop and validate a method for the generation and insertion of digitally synthesized polyps into CT images of the human colon for use as ground truth in evaluations of virtual colonoscopy. In this article, we describe and evaluate our method, which allows generation of realistic polyps of various sizes and shapes that can be placed anywhere inside a patient's CT-imaged colon. This high degree of control over polyp characteristics allows the generation of multiple phantom CT data sets, which can then be used for systematic evaluation of the trade-offs among the many display techniques currently under investigation for CT colonography. Such an approach can help determine which display method, if any, will prove to be a feasible clinical tool.
|
MATERIALS AND METHODS |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
CT Colonographic Acquisition
For this study, we selected CT colonographic data sets obtained in eight patients (five men, three women; mean age, 56 years; age range, 30–75 years); the data sets were part of a separate ongoing clinical trial (4), approved by our institutional review board, to compare CT colonography and fiberoptic colonoscopy. In all participants, previous sigmoidoscopic results were positive, and the participants had been referred for fiberoptic colonoscopy.
Prior to CT, patients underwent standard precolonoscopy cleansing with a solution of polyethylene glycol and electrolytes (GoLYTELY; Braintree Laboratories, Braintree, Mass). After informed consent had been obtained, CT was performed (CT HiSpeed Advantage scanner; GE Medical Systems, Milwaukee, Wis) at 120 kVp, 200 mA, 3-mm collimation, and 6 mm/sec table speed (pitch, 2.0). The patient was in the supine position, and air insufflation of the colon was performed. Patients were instructed to hold their breath as long as comfortably possible, with quiet breathing, if necessary, during the later part of the study, which covered the pelvic region. CT scanning extended from the dome of the diaphragm to the pubic symphysis, with a total acquisition time of 60 seconds.
Scans were reconstructed at 1-mm intervals with a display field of view of 36–40 cm by using the manufacturer's standard soft-tissue reconstruction algorithm. All images were reconstructed with a 512 x 512 matrix, which resulted in an in-plane pixel size of 0.625–0.703 mm. Reconstructed image data sets were transferred to the Stanford University Department of Radiology 3D Medical Imaging Laboratory.
A positive CT finding for at least one polyp was obtained in each of the eight patients included in this study, and the CT finding was directly correlated with a positive finding at fiberoptic colonoscopy. We selected imaging data from one of the eight patients (a 60-year-old man), hereafter referred to as the "base" data set, for synthetic polyp insertion. For comparison, we chose 10 regions that contained a single true polyp from the data sets obtained in the other seven patients. These 10 polyps were sessile, with various shapes and sizes between 8 and 13 mm.
Creation of Phantom Data
All software was implemented on workstations (O2; Silicon Graphics, Mountain View, Calif) equipped with an R10000 processor and 256 Mbyte of main memory. All volume rendering was performed by using VOXELVIEW software (version 2.5.4; Vital Images, Minneapolis, Minn).
Polyp synthesis.—To create the synthetic polyps, we used a software spiral CT simulator. This simulator allows one to define the shape of an object with geometric primitives and generates projection data and CT reconstructions by using configurable scanner geometry, acquisition parameters, and reconstruction parameters. We configured the simulator parameters to match those of the CT scanner used to acquire the patient data (Table). Because polyps are small, however, we needed to simulate only those projections that covered a small field of view (22.5 mm) and to reconstruct the images with a 32 x 32 matrix. The resultant in-plane pixel size (0.703 x 0.703 mm) matched that of the base data set we chose for synthetic polyp insertion. In addition, the simulator allowed specification of the variance of a Poisson noise source to be added to the detector measurements, thereby correlating the noise in the reconstructed images in a manner similar to that in reconstructions obtained with actual CT scanners. We adjusted the Poisson noise variance empirically, so that the SDs in regions of interest in the polyps were in the same range as similar measurements made in soft tissue in the base data (approximately 15–25 HU).
|
The shape of each polyp was defined by using a combination of geometric primitives, such as ellipsoids, cylinders, and cones. By using combinations of such primitives, lesions of nearly any shape could be constructed, the simplest being a single sphere. Lesions with a cylindric base and a hemispheric cap were constructed by simulating scans of an object represented by the union of a cylinder and a sphere. Nodular, grossly hemispheric lesions were constructed by simulating scans of an object represented by the union of one large sphere (approximately 10 mm in diameter) and multiple smaller spheres (2–3 mm in diameter) arranged around the surface of the large sphere. We simulated the CT scans of each user-defined polyp and reconstructed 30 cross sections (to cover 30 mm) through each lesion.
Each voxel in the reconstructed volume contained a 16-bit integer value that corresponded to the CT intensity (in Hounsfield units) of the scanned polyp. The polyp was inserted into the patient's colon data by using the algorithm described in the next paragraph.
Polyp insertion.—We developed a software package that allows the user to insert a synthetic polyp into a volume CT study. The key features of the insertion algorithm are that it allows (a) smooth insertion of the synthetic polyp into patient data and (b) maintenance of the partial volume effects and noise in the synthetic polyp and original patient data. The insertion algorithm operates on a voxel-by-voxel basis by using either direct voxel replacement or nonlinear addition with no spatial averaging.
To understand the insertion algorithm, consider a fixed location in the original patient data with a voxel CT intensity IP. At the corresponding location in the synthetic polyp data set, there will be a voxel CT intensity IS. The insertion algorithm combines IP and IS into a final voxel intensity IF. There are two cases to consider, depending on whether the patient voxel definitely lies inside soft tissue. When IP > 0.95 x IA, where IA is the average CT intensity of the tissue surrounding the air-filled lumen, we can be fairly certain that the voxel lies within tissue; in this case, we set IF = IP. When IP 
0.95 x IA, we assume that the voxel is fully or partially in the air-filled lumen; in this case we set IF = IS + IP - IS x IP/IA.
This method, known as compositing, is commonly used in computer graphics for smoothly blending two images together (14). Note that when IS and IP are small, IF is approximately equal to the sum of IS and IP. When IS and/or IP is nearly equal to IA, IF is saturated at the expected value IA. This saturation, though, leads to a substantial reduction in the noise in those voxels for which both IS and IP are large. We therefore compensated for this effect by adding spatially correlated noise weighted by the factor [2 - (1 - IP/IA)2 - (1 - IS/IA)2]
. Note that when IS and IP are small, the weight is small, and little noise is added. The compensation increases, however, as IS and IP approach IA. The correlated noise source was created by simulating a large spherical object with the CT simulator.
Quantification of polyp edge profiles and comparison with features of the base CT data were performed by using software (NIH IMAGE; National Institutes of Health. Available at: http://rsb.info.nih.gov/nih-image/. Accessed January 10, 1998.) and various models of Macintosh desktop computers (Apple Computers, Cupertino, Calif). Profiles of image intensity versus pixel number were generated in both the in-plane and the through-plane directions. Through-plane intensity profiles were generated from sagittal or coronal reformatted images created with VOXELVIEW. Note that the intensity profile obtained at the traversal from air to soft tissue, such as that encountered at the normal colonic wall, should be analogous to the profile obtained at the traversal from air to a polyp with soft-tissue attenuation. Intensity profiles with which synthetic lesions were compared with features of the native colon were inspected visually, and iterative adjustments were made in the simulation parameters to minimize observed differences between the profiles.
To determine how the polyp synthesis and insertion techniques work in a simple case, consider a 10-mm spherical polyp that was generated as described earlier. Figure 1 shows representative in-plane and longitudinal profiles for the air-tissue interfaces in both a patient colon and a synthetic polyp. This particular polyp required 30 seconds of computation time to be generated. (More complex shapes may take up to 2 minutes.) One can clearly see that the profiles of the air-tissue and air-polyp interfaces agreed well in both the in-plane and the longitudinal directions, with the 10%–90% transition occurring over 2.1 pixels (1.5 mm) in-plane, and 3.7 pixels (3.7 mm) longitudinally. These results show that partial volume effects can be accurately simulated with our system, and the same degree of edge blurring in synthetic lesions as in native tissue-air interfaces in the base colon data can be provided. This modeling is necessary to eliminate spurious visual cues such as overly sharp edges when viewing simulated lesions in polyp detection trials.
|
|
|
|
|
We generated 3D virtual endoscopic volume-rendered images of the synthetic polyps and the real polyps by using VOXELVIEW. The color, opacity, and lighting settings for all polyps were similar to those reported by McFarland et al (12). We saved 2D color images of the renderings, each annotated with a black caret that clearly identified each polyp.
In addition to the 3D volume-rendered images, standard gray-scale 2D axial CT images of each polyp were obtained. The level and window settings were -200 and 800 HU, respectively. In addition, each image was magnified by a factor of three to render it more visible. Each gray-scale image was then annotated with a black caret to identify the polyp.
The 40 images (20 volume-rendered and 20 axial CT images) were then incorporated onto a single World Wide Web page. All images were stored in Joint Photographic Experts Group, or JPEG, format, with the quality factor set to the highest level (ie, with minimal compression). We arranged the 3D and 2D images of the 40 polyps on the page in random order.
Three radiologists (including R.B.J.) were asked to independently score each image by using the following World Wide Web–based interface. Next to each image was a series of five "radio buttons." Button 1 was for rating a lesion as definitely synthetic; button 2, for rating a lesion as possibly synthetic; button 3, for a rating of unsure if synthetic or real; button 4, for rating a lesion as possibly real; and button 5, for rating a lesion as definitely real. Only one button could be toggled for each image. The readers were not told the number of real and synthetic polyps. They were told only that there was a mixture of real and synthetic polyps and that there was no correlation between the 2D axial sections and the 3D volume-rendered images. Each reader was asked to spend approximately 20 seconds studying each image, but no time limit was imposed.
Results for each reader were analyzed as frequency distributions that represented the number of responses in each of the five categories. The frequency histograms for individual readers were not significantly different from each other, so we pooled the results to maximize the statistical power of the observations and as a convenient means of data summary.
Two analyses were performed with the results to test for differences between real and synthetic polyps. First, the frequency distributions for real and synthetic polyps were compared against each another for each of the 3D and 2D display modes. The Wilcoxon rank sum test was used, which is appropriate for ordinal data without the need to assume an underlying normal distribution, with a P value of .05 or less indicative of a significant difference (15).
Second, receiver operating characteristic (ROC) curve analysis was performed by using the program ROCKIT (version 0.9B [beta]; Metz CE, Chicago, Ill). In this analysis, the input data were pairs of values that represented each category along the spectrum from "definitely synthetic" to "definitely real" and its corresponding response frequency for the pooled readers' results. By varying the definition of a true-positive response across the categories, pairs of ROC operating points that represented true-positive observations (sensitivity) versus false-positive observations (1 - specificity) were generated. Mathematically fitted ROC curves were compared against an ROC curve that represented random guesses about which polyps were real and which were synthetic. (This "random guess" curve is a diagonal line, and the area under the curve, or Az value, is 0.5). The Az values, 95% CIs, SDs of Az values, and P values for ROC curve comparison were generated by the ROCKIT program.
|
RESULTS |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
|
|
|
For the 2D axial CT display (Fig 6b), there also was considerable overlap between the frequency distributions. For real lesions, the median and modal scores were 3 and 2, respectively, whereas for the synthetic polyps, the median and modal scores were both 3. The test for statistically significant differences between real and synthetic polyps yielded a P value of .11, which again was indicative of no difference. Note that although no significant differences between the two distributions were detected, there was a tendency with 2D axial CT sections for readers to interpret real lesions as synthetic.
By using ROC curve analysis, we evaluated whether the readers performed significantly better than would be expected from random guessing. Figure 7 shows ROC curves that depict the true-positive fraction (sensitivity) as a function of the false-positive fraction (1 - specificity). An observed ROC curve that is positioned above and to the left of the diagonal line that represents random guessing would be suggestive of the ability to distinguish real from synthetic lesions. For 3D displays, the ROC curve closely paralleled that of random guessing, with an Az value of 0.54 (SD = 0.08; 95% CI = 0.39, 0.68). The test for significant differences between this curve and random guessing yielded a P value of .70, or no significant difference. For 2D axial CT, the ROC curve tended to be positioned below the level expected for random guessing, with an Az value of 0.39 (SD = 0.07; 95% CI = 0.25, 0.53). The test for significant differences between this curve and random guessing yielded a P value of .28, or no significant difference.
|
|
DISCUSSION |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
To create synthetic polyps that appear to be composed of real tissue, the synthetic polyps must have the correct partial volume effects. Mismatched partial volume effects can lead to artifacts on both 2D axial sections and 3D volume-rendered images; these artifacts can cause the synthetic polyps to appear conspicuously "computer generated" rather than resemble real tissue. Thus, we chose to create polyps with a CT simulator to match the partial volume effects in synthetic polyps with those in real tissue.
A second important factor is that the noise in the synthetic polyp must match the noise in the tissue. The CT simulator permits addition of noise to the simulated projections, thereby resulting in synthetic polyps that have noise fluctuations with SD and spatial correlation parameters similar to those seen in tissue.
Another important factor is that the polyps must have a realistic shape. Naturally occurring sessile polyps come in a variety of types, such as marble, mountain, clam shell, and carpet (16). To our knowledge, the frequency of occurrence of specific polyp types in any given patient population has not been reported. Nevertheless, our technique provides the flexibility to generate almost any type of shape by using a combination of ellipsoids, cones, and cylinders.
To validate that the polyps we generated had a realistic appearance, we performed a study to determine if radiologists could distinguish synthetic from real polyps. Because the artifacts seen on 3D images may not be the same as those seen on 2D images, we performed separate evaluations for 3D volume-rendered images and 2D axial sections (Fig 6). Whether we analyzed the results to detect for significant differences between the histograms of radiologists' scores for real and synthetic lesions or the ROC curves to compare observer performance with the hypothetical case of random guessing, we did not detect statistically significant differences between real and synthetic lesions. In fact, there was a tendency (nonsignificant) for readers to interpret real lesions as synthetic on 2D axial CT sections (Fig 6b), a factor that resulted in the ROC curve for the 2D display to be positioned somewhat below the curve for random guessing (Fig 7).
Precisely what visual features of lesions were underlying these trends is unknown; however, a bias toward interpretation of real lesions as synthetic is the opposite of would be expected if the readers could truly identify synthetic lesions. This provides some reassurance that our methods for creation and insertion of synthetic lesions are sound. On the basis of results from the validation study, we conclude that the synthetic lesions mimic real lesions closely enough to be used as a model system for tests of display modes in virtual colonoscopy.
Other groups have developed different methods to construct colonic polyp phantoms. One technique involves the use of a pig colon (17), in which polyps are formed by "puckering" the mucosa with suture material. A limitation of this method is that the haustral folds in a pig colon are different from those in a human colon, and human colonic polyps are often located on or behind a fold. Also, a limited number of polyps may be formed in a single specimen, and the precision and range of shapes available are limited. The use of ex vivo specimens also makes it difficult to mimic the overall complexity involved in a study of the entire human colon.
Another phantom-generation method involves direct copying of voxels from images of patient tissue and insertion of those voxels near the imaged colonic surface (2). However, this method cannot be used to achieve partial volume effects at the edge of the simulated polyp that are the same as those in the nearby colonic wall, and the spatial discontinuity in attenuation at the edge of the "pasted" polyp may render it more visible than a real polyp, thereby biasing the results for display and/or detection methods.
Our approach is most closely related to that reported by Seltzer et al (18), who analytically calculated the partial volume effects for spherical lung nodules and volume averaged the analytic models on a voxel grid matched to patient CT scans of the lungs. In the lung, extension of these methods to other than spherical shapes is not necessary, and the problem of insertion of synthetic nodules at the edge of a structure with comparable attenuation was not addressed. Finally, we note that an important feature of our technique is that any number of user-defined polyps can be placed into patient data at random locations; when used in evaluation of display and detection methods, review of these data sets closely mimics the overall complexity and tedium of review of actual patient studies.
One limitation of our method was that the creation of different polyp shapes is more an art than an exact science. Each polyp was designed by hand, one at a time, with random asymmetries. Because polyps should not be identical, this process had to be repeated for each new polyp. A more systematic approach would be to prepare a library of basic polyp shapes from which the user could choose. The user could then further specify the size and orientation of the polyp. The program would generate a list of ellipsoids, cones, and cylinders that would constitute the polyp. In addition, the program could make small random modifications to the shape so that each polyp is unique.
Another potential limitation of our study was our use of one base data set obtained in a single patient, rather than several data sets from different patients, for insertion of synthetic polyps. However, because the geometry of even a single colon varies considerably over its length and because we inserted lesions at numerous locations in the colon, the only variability one would expect among patient data sets would result from differences in scanning parameters and image noise levels. Note that in the early stages of this work and later in subsequent studies, we did use imaging studies obtained in several patients as base data. Our experience suggests that as long as individual variations in pixel dimensions and image noise are considered, the simulation and insertion process appears to be valid.
A final limitation of our study was that although we could not demonstrate a significant difference between observations of real and synthetic lesions, this result does not explicitly prove that simulated lesions are identical to real lesions. This lack of a difference was largely a result of the relatively small number of real lesions with which the synthetic lesions could be compared. It is no coincidence that this paucity of real lesions provided much of the motivation for the development of synthetic lesions in the first place. In the context of the current work, we can conclude that the synthetic lesions did not contain notably "artificial" features that allowed their differentiation from real lesions. From this, one can make the assumption, albeit untested, that the conspicuity of synthetic lesions should be similar to that of real lesions. Given these limitations, we believe that our results are sufficiently sound to begin application of the phantom system to initial trials in which the display modes are compared (19).
In summary, our polyp synthesis and insertion method allows the generation of a variety of realistic colonic polyps that could not be distinguished from real lesions by experienced observers and, therefore, can facilitate the comparison of display techniques. The fine control over polyp shape, size, and location will facilitate the development and optimization of the many display tools currently under investigation for CT colonography.
|
Acknowledgments |
|---|
|
Footnotes |
|---|
Abbreviations: ROC = receiver operating characteristic 2D = two-dimensional 3D = three-dimensional
Author contributions: Guarantors of integrity of entire study, C.F.B., S.N.; study concepts and design, C.F.B., C.K.; definition of intellectual content, C.F.B., C.K., S.N.; literature research, C.F.B., C.K.; experimental studies, C.F.B., R.B.J., C.K.; data acquisition, D.S.P., C.K.; data analysis, C.K., C.F.B.; statistical analysis, C.F.B., C.K.; manuscript preparation, C.K., C.F.B., S.N.; manuscript editing, R.B.J., D.S.P.; manuscript review, all authors.
|
References |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
Related Article
This article has been cited by other articles:
|
|
 |
|
  T. G. Mang, C. Schaefer-Prokop, A. Maier, E. Schober, G. Lechner, and M. Prokop Detectability of Small and Flat Polyps in MDCT Colonography Using 2D and 3D Imaging Tools: Results from a Phantom Study Am. J. Roentgenol., December 1, 2005; 185(6): 1582 - 1589. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
A. Ozgun, E. Rollven, L. Blomqvist, S. Bremmer, R. Odh, and A. Fransson Polyp Detection with MDCT: A Phantom-Based Evaluation of the Impact of Dose and Spatial Resolution Am. J. Roentgenol., April 1, 2005; 184(4): 1181 - 1188. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. Yee, N. N. Kumar, R. K. Hung, G. A. Akerkar, P. R. G. Kumar, and S. D. Wall Comparison of Supine and Prone Scanning Separately and in Combination at CT Colonography Radiology, March 1, 2003; 226(3): 653 - 661. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. M. Summers, C. F. Beaulieu, L. M. Pusanik, J. D. Malley, R. B. Jeffrey Jr, D. I. Glazer, and S. Napel Automated Polyp Detector for CT Colonography: Feasibility Study Radiology, July 1, 2000; 216(1): 284 - 290. [Abstract] [Full Text]
|
|
|
|
|
|
C. F. Beaulieu, R. B. Jeffrey Jr, C. Karadi, D. S. Paik, and S. Napel Display Modes for CT Colonography: Part II. Blinded Comparison of Axial CT and Virtual Endoscopic and Panoramic Endoscopic Volume-rendered Studies Radiology, July 1, 1999; 212(1): 203 - 212. [Abstract] [Full Text]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
) and the edge of patient's colon (
) as a function of pixel location (X in a, Z in b). As one moves from the lumen (air attenuation, -1,000 HU) to the polyp or colonic wall (soft-tissue attenuation, approximately 50 HU), the slope of the line describing the change in attenuation is determined by either in-plane or through-plane partial volume averaging. By matching the input parameters of the CT simulator with those of the base data, closely matched profiles can be obtained.
). If the readers had been able to distinguish real from synthetic lesions, the ROC curves for 3D and/or 2D images would be positioned above and to the left of the curve that represents random guessing (dashed line). Instead, the observed ROC curves were not significantly different from random guessing. FPF = false-positive fraction, TPF = true-positive fraction.