Published online before print December 21, 2005, 10.1148/radiol.2381050477
(Radiology 2005;238:699.)
A more recent version of this article appeared on December 1, 2005
© RSNA, 2005
Simulation of Liver Lesions for Pediatric CT1
Chee L. Hoe, MS,
Ehsan Samei, PhD,
Donald P. Frush, MD and
David M. Delong, PhD
1 From the Duke Advanced Imaging Laboratories, Department of Radiology (C.L.H., E.S.), and Division of Pediatric Radiology, Department of Radiology (D.P.F.), Duke University Medical Center, Box 3302, Durham, NC 27710; and Departments of Physics (C.L.H., E.S.), Biomedical Engineering (E.S.), and Biostatistics and Bioinformatics (D.M.D.), Duke University, Durham, NC. From the 2004 RSNA Annual Meeting. Received March 22, 2005; revision requested May 12; revision received June 6; final version accepted June 24. Supported in part by a grant from GE Healthcare.
Address correspondence to E.S. (e-mail: samei{at}duke.edu).
 |
ABSTRACT
|
|---|
Purpose: To develop and validate a technique based on characteristics of real lesions for simulating realistic small liver lesions on pediatric computed tomographic (CT) images.
Materials and Methods: The institutional review board provided exempt status for this study, determined that it was not subject to HIPAA compliance, and did not require informed consent. Patient identification information was removed from clinical images from contrast material–enhanced multi–detector row CT examinations performed in 10 children. Patients were infants or children up to 18 years old. Information about sex was not available. Children had one or more liver lesions of 2–6 mm in maximum transverse diameter. Images with more than one lesion were rendered multiple times, and each time, all but one of the lesions were digitally removed in sequence. This process provided images (n = 19) with a single real lesion. For consistency, the same image backgrounds (images with all real lesions removed) were used to create an identical number of images (n = 19), each with a single simulated lesion. Subsequently, three radiologists independently assessed images of real and simulated lesions that were presented in random order with a score on a continuous scale of 0 (definitely simulated) to 100 (definitely real). Mixed-model analysis of variance was used to test the null hypothesis that the difference in population mean scores between the two lesion types was zero.
Results: The observer study did not reveal a significant difference in the ability of any radiologist to discriminate between real and simulated lesions (P > .31). The differences in mean scores for discrimination between real and simulated lesions for the three observers were –6, 9, and –7, respectively. The estimated overall difference was –1.
Conclusion: Mathematic simulation of liver lesions is a feasible technique for creating realistic lesions for image quality or dose reduction studies in pediatric CT.
© RSNA, 2005
|
INTRODUCTION
|
|---|
Since its inception by Hounsfield in 1972, computed tomography (CT) has been established as an essential diagnostic modality for use in both adults and children (1,2). Compared with other medical imaging modalities, CT delivers a higher radiation dose to patients. Children are more susceptible to this radiation dose because of the prolonged potential for cancer induction (3). The number of pediatric CT examinations is not inconsequential, as pediatric CT accounts for up to 11% of all CT examinations (4), which is estimated to be more than 7 000 000 examinations per year in the United States alone (5). While there has been a call to reduce radiation dose in pediatric body CT examinations (6), there also has been a need for systematic evaluation of the effect of dose reduction on image quality and diagnostic capability (7). The balance between image quality and radiation dose reduction is not clear, especially for subtle but clinically important lesions. Investigation of this balance is challenging because of ethical considerations (eg, radiation exposure from repeated sequential CT examinations at different dose levels in a child), methodological issues (eg, changes in liver lesion conspicuity during the contrast enhancement process), and very low occurrence of subtle single lesions in children.
To circumvent these problems, techniques for simulation of lesions can be very valuable. In the past, two simulation techniques have been used: (a) the embedment of round objects in a phantom (8) and (b) the addition of computer-simulated ellipsoidal objects to existing CT images of patients (9–11). These techniques have disadvantages: They do not create lesions on clinical images (technique a), and/or they do not create realistic lesions (techniques a and b). The purpose of our investigation, therefore, was to develop and validate a technique based on characteristics of real lesions for simulating realistic small liver lesions on pediatric CT images.
|
MATERIALS AND METHODS
|
|---|
This study was supported in part by a grant from GE Healthcare, Waukesha, Wis; however, the authors had complete control of the data and information submitted in this article.
Image Selection
Images from which patient identification information was removed (teaching file) were used. These images were from examinations of the liver with contrast material–enhanced multi–detector row CT (LightSpeed; GE Healthcare) performed in 10 patients who were infants or children up to 18 years old. Information about sex was not available. For selected images, one or more small liver lesions were identified by one of the authors (D.P.F.), who had 13 years of experience with pediatric CT. Lesions represented microabscesses (n = 4) or hepatic involvement from systemic disorders (Langerhans cell histiocytosis, n = 1; lymphoma, n = 1; or metastatic disease, n = 4). Lesions were categorized on the basis of the presence of cultures from other sites, such as blood, that were positive for organisms and/or a response to antibiotics, such as that in microabscesses, or the presence of systemic disease for which the clinical diagnosis was hepatic involvement (Langerhans cell histiocytosis, lymphoma, and malignancy). No biopsies of lesions were performed. Our institutional review board provided exempt status for this study, determined that it was not subject to compliance with the Health Insurance Portability and Accountability Act, and did not require informed consent.
Lesion Characterization and Simulation
The size and attenuation difference of the lesions (ie, the difference in Hounsfield units between the lesions and the background, also referred to here as lesion contrast) in the selected images were then tabulated by two authors (D.P.F. and C.L.H.). The lesion size was 2–6 mm in diameter, and the lesion contrast ranged from –20 to –41 HU. Images with lesions greater than 6 mm or with excessive noise (ascertained visually) were excluded.
From the selected images, square regions of interest (ROIs) with a lesion occupying approximately half of the ROI were identified (Fig 1a). The background removal process is shown in Figure 1b–1e. The ROIs were first segmented (Fig 1b), followed by removing the lesion and filling the gap by interpolating the surrounding region (Fig 1c and 1d). The interpolation was performed by using a digital in-painting technique (John D'Errico, MS, written communication, March 27, 2004).
View larger version (100K):
[in this window]
[in a new window]
|
Figure 1a: Background removal process and lesion simulation. (a) CT image with ROI that includes a lesion, (b) the segmented ROI, (c) the ROI with the lesion and surrounding area removed, (d) the ROI with the lesion background interpolated, (e) the segmented lesion profile (obtained from subtracting d from b), (f) a mask of the lesion profile fitted according to an analytic expression, and (g) a simulated lesion created from subtracting the fitted mask in f from a CT image. All of the ROIs were magnified 12 times. Images are shown in transverse view.
|
|
View larger version (109K):
[in this window]
[in a new window]
|
Figure 1b: Background removal process and lesion simulation. (a) CT image with ROI that includes a lesion, (b) the segmented ROI, (c) the ROI with the lesion and surrounding area removed, (d) the ROI with the lesion background interpolated, (e) the segmented lesion profile (obtained from subtracting d from b), (f) a mask of the lesion profile fitted according to an analytic expression, and (g) a simulated lesion created from subtracting the fitted mask in f from a CT image. All of the ROIs were magnified 12 times. Images are shown in transverse view.
|
|
View larger version (96K):
[in this window]
[in a new window]
|
Figure 1c: Background removal process and lesion simulation. (a) CT image with ROI that includes a lesion, (b) the segmented ROI, (c) the ROI with the lesion and surrounding area removed, (d) the ROI with the lesion background interpolated, (e) the segmented lesion profile (obtained from subtracting d from b), (f) a mask of the lesion profile fitted according to an analytic expression, and (g) a simulated lesion created from subtracting the fitted mask in f from a CT image. All of the ROIs were magnified 12 times. Images are shown in transverse view.
|
|
View larger version (101K):
[in this window]
[in a new window]
|
Figure 1d: Background removal process and lesion simulation. (a) CT image with ROI that includes a lesion, (b) the segmented ROI, (c) the ROI with the lesion and surrounding area removed, (d) the ROI with the lesion background interpolated, (e) the segmented lesion profile (obtained from subtracting d from b), (f) a mask of the lesion profile fitted according to an analytic expression, and (g) a simulated lesion created from subtracting the fitted mask in f from a CT image. All of the ROIs were magnified 12 times. Images are shown in transverse view.
|
|
View larger version (56K):
[in this window]
[in a new window]
|
Figure 1e: Background removal process and lesion simulation. (a) CT image with ROI that includes a lesion, (b) the segmented ROI, (c) the ROI with the lesion and surrounding area removed, (d) the ROI with the lesion background interpolated, (e) the segmented lesion profile (obtained from subtracting d from b), (f) a mask of the lesion profile fitted according to an analytic expression, and (g) a simulated lesion created from subtracting the fitted mask in f from a CT image. All of the ROIs were magnified 12 times. Images are shown in transverse view.
|
|
The term in-painting stems from its use by art conservators, who fill in missing parts in precious artworks by using cues from surrounding areas. The digital in-painting technique is the digital counterpart of the in-painting technique in art. It essentially is an interpolation algorithm that solves a partial differential equation in an unknown region based on data from surrounding regions. The digital in-painting technique enables formulation of a partial differential equation that is assumed to apply over all regions. Where there are data, the partial differential equation is ignored; in the unknown regions, the partial differential equation is solved as a boundary value problem. In this work, the partial differential equation applied in the unknown regions was approximated with a finite difference approximation. This resulted in a linear system of equations, which can be solved by using linear algebra. To make the solution more efficient, sparse matrices were used. There are a few candidates for the partial differential equation in the in-painting technique. The Laplace equation was used in this work (12).
The interpolated image was subtracted from the original image, and with this process, a mask of the lesion remained that was relatively free of the background structure (Fig 1e). The mask was then fitted with the radially symmetric lesion model of Samei et al (13,14) and Burgess et al (15), which is calculated as c(r) = C (1 – D r2)n, with exponents of 1.0, 1.5, and 2.0. In this model, the lesion contrast, or c(r), of a particular pixel within the lesion varies as a function of its distance from the center of the lesion, called r. The two constants C and D determine the peak lesion contrast and size of the lesion, respectively. This equation was suggested by Samei et al in a slightly different form (13,14). Burgess et al (15) suggested the current form that is conducive to Fourier transformation. A radiographic representation of this equation for the exponent 1.5 is depicted in Figure 1f.
View larger version (53K):
[in this window]
[in a new window]
|
Figure 1f: Background removal process and lesion simulation. (a) CT image with ROI that includes a lesion, (b) the segmented ROI, (c) the ROI with the lesion and surrounding area removed, (d) the ROI with the lesion background interpolated, (e) the segmented lesion profile (obtained from subtracting d from b), (f) a mask of the lesion profile fitted according to an analytic expression, and (g) a simulated lesion created from subtracting the fitted mask in f from a CT image. All of the ROIs were magnified 12 times. Images are shown in transverse view.
|
|
The analytic representation was the model used for simulation of lesions. When the analytic representation is subtracted from normal liver multi–detector row CT images, it could create simulated lesions with a radiographic appearance similar to those of real lesions (Fig 1g). All computer codes used in this study were written in an interactive computer program (Matlab; Mathworks, Natick, Mass).
View larger version (101K):
[in this window]
[in a new window]
|
Figure 1g: Background removal process and lesion simulation. (a) CT image with ROI that includes a lesion, (b) the segmented ROI, (c) the ROI with the lesion and surrounding area removed, (d) the ROI with the lesion background interpolated, (e) the segmented lesion profile (obtained from subtracting d from b), (f) a mask of the lesion profile fitted according to an analytic expression, and (g) a simulated lesion created from subtracting the fitted mask in f from a CT image. All of the ROIs were magnified 12 times. Images are shown in transverse view.
|
|
Pilot Validation
As an initial assessment of the quality of simulation, an experienced pediatric radiologist (D.P.F.) evaluated a set of eight images with two real and six simulated lesions. The radiologist was first asked if he could distinguish the real lesions from the simulated lesions. He was then asked which of the three exponents produced lesions that best mimicked real lesions. Subjectively, all the lesions generated through the three different exponents fit well with profiles of the real lesions in our database; however, the closest visual match was found when the lesions were generated by using the analytic expression with an exponent of 1.5. Use of this exponent is consistent with that in the earlier work of Samei et al (13,14) for simulating lung nodules on chest radiographs, and this exponent was used in this study.
With the size and contrast of real lesions as a guide, a range of size and contrast for simulated lesions was identified. To ensure that the ranges for size and contrast were appropriate, lesions that demonstrated the extremes of the lesion contrast and size were simulated on example images (Fig 2). These images were examined subjectively by an experienced pediatric radiologist (D.P.F.). Contrast of simulated lesions was changed until a range that represented the range of pixel characteristics of real lesions (–20 to –41 HU) was obtained.
View larger version (104K):
[in this window]
[in a new window]
|
Figure 2: An example transverse CT image for pilot determination of lesion size and contrast of simulated lesions (arrows).
|
|
Comparison of Real and Simulated Lesions
Given the small size of our database (10 patients), images with more than one lesion were used multiple times to create a larger number of images, each with a single real lesion. For example, an image with three small lesions would be used to create three images, each with a single lesion. These images were created by one of the authors (C.L.H.). To do so, images with more than one lesion were rendered multiple times, and each time, all but one lesion was digitally removed in sequence. To make the eventual images look real, a square ROI that contained the lesion was first extracted. That region was then filled in with a slightly smaller ROI that was copied from a region adjacent to the extracted ROI. Finally, the same digital in-painting method described earlier was used to interpolate the remaining region surrounding the filled-in region. In this manner, 10 initial multi–detector row CT images (with one or more real lesions) were used to create 19 images with a single real lesion. An example of this lesion isolation process is shown in Figure 3.
View larger version (103K):
[in this window]
[in a new window]
|
Figure 3a: Transverse CT image with (a) multiple real lesions (arrows) and with (b) an isolated real lesion (arrow). This step was for creation of a large number of images that each included a single real lesion for the observer study.
|
|
View larger version (103K):
[in this window]
[in a new window]
|
Figure 3b: Transverse CT image with (a) multiple real lesions (arrows) and with (b) an isolated real lesion (arrow). This step was for creation of a large number of images that each included a single real lesion for the observer study.
|
|
After identification of the suitable range for both lesion contrast and size, the ranges were randomly sampled in creating 19 images with simulated lesions. For consistency, these 19 images were created from the same 10 initial multi–detector row CT images. All lesions from these 10 images were removed by using the method described earlier before insertion of the simulated lesions. Counting the usage for both real and simulated images in the full set of 38 images, eight of the 10 initial multi–detector row CT images were used four times and two of them were used three times. Figure 4 shows one of the final 19 images with a simulated lesion.
View larger version (103K):
[in this window]
[in a new window]
|
Figure 4: Transverse CT image with a simulated lesion (arrow). This was one of the images used in the observer study.
|
|
Observer Study
The final step was a blinded observer study. All 38 images (19 images with a single real lesion and 19 images with a single simulated lesion) were loaded onto a display workstation with a calibrated 5-megapixel high-fidelity cathode ray–tube display (MGD 521; Barco, Kortrijk, Belgium). Three experienced pediatric radiologists (with 1, 3, and 18 years of experience with pediatric CT) in an identical controlled lighted environment used a graphical user interface that enabled them to rate the lesions on a confidence scale from 0 (definitely simulated) to 100 (definitely real). This graphical user interface was modified from a graphical user interface for the study of breast lesions (16). The order in which the images were shown was randomized for each observer. All the observers were told there was one and only one lesion on each image; they were not told how many of the lesions were real.
Lesions were rated on the basis of how real they looked to each observer. Observers based their decision on experience garnered through training and clinical service in pediatric radiology at a tertiary care university-based Division of Pediatric Radiology. Observers were encouraged to use the full scale for assigning scores, and they had no time limit in going through the images. To become familiarized with the graphical user interface, each observer participated in a brief trial assessment with a separate set of 16 images before assessment of the test set of 38 images. Observers had the option of adjusting the window width and window level for each image, if desired. Each observer was asked to point to the lesion by using a mouse before rating it to assure that the rating was not performed for a spurious lesion. All scores were automatically transferred to a text file.
Data Analysis
A mixed-model analysis of variance in which reader and image effects were treated as random was used to test the null hypothesis that the difference in mean scores between real and simulated lesions would be zero. The variance components from the mixed model were used to estimate an interreader correlation coefficient (17,18). Kendall 
correlation coefficients (19) also were used to assess interreader agreement. These coefficients were used to estimate the probability that the observers would agree on the relative assignment of a score for two lesions, that is, if the first lesion was assigned a higher score than the second lesion. The Kendall correlation coefficients were calculated for both real and simulated lesions. Descriptive statistics and t statistics were computed for scores assigned to both types of lesions for all three observers. These statistical analyses were performed with commercial statistical software (SAS, version 8.2; SAS Institute, Cary, NC) by using a 95% confidence level; a difference with a P value of .05 was considered statistically significant.
As a graphical depiction of the quality of simulated lesions, a histogram of the frequency of score assignment versus the score assigned for both types of lesions was derived for each observer. A summary histogram for all observers was also obtained. In all histograms, scores smaller than N but bigger than (N – 10) were binned under N, for N = 10–100.
|
RESULTS
|
|---|
The estimated difference in mean scores between real and simulated lesions from the analysis-of-variance model averaged across observers was –1.3 (estimated standard error of the mean difference, 5.3; P > .81). The corresponding intraclass correlation estimate was 0.08, and this value indicated little agreement among the readers on the assignment of scores to the lesions. The range of Kendall correlation coefficients was 0.25–0.33. Statistically, these coefficients were not significantly different from zero. The difference between the mean score of real lesions and that of the simulated lesions for the three observers ranged from –7 to 9 of a maximum of 100, with individual P values of .52, .35, and .31 for observers 1, 2, and 3, respectively. The sample standard deviations of scores for the three readers varied from 21 to 30, and the maximum value for standard error was 9.4. The differences between the maximum and minimum scores for the real and simulated lesions for all observers were small (Table).
The histogram of the scores (Fig 5) showed that the scores were evenly distributed for both lesion types for all observers. Scores assigned by observers 1 and 2 were slightly more spread out than those assigned by observer 3.
View larger version (21K):
[in this window]
[in a new window]
|
Figure 5a: Histogram shows frequency of score assignment versus score for (a) observer 1, (b) observer 2, (c) observer 3, and (d) all observers. Binned score for value N refers to the observer scores greater than (N – 10) and smaller than N.
|
|
View larger version (20K):
[in this window]
[in a new window]
|
Figure 5b: Histogram shows frequency of score assignment versus score for (a) observer 1, (b) observer 2, (c) observer 3, and (d) all observers. Binned score for value N refers to the observer scores greater than (N – 10) and smaller than N.
|
|
View larger version (19K):
[in this window]
[in a new window]
|
Figure 5c: Histogram shows frequency of score assignment versus score for (a) observer 1, (b) observer 2, (c) observer 3, and (d) all observers. Binned score for value N refers to the observer scores greater than (N – 10) and smaller than N.
|
|
View larger version (23K):
[in this window]
[in a new window]
|
Figure 5d: Histogram shows frequency of score assignment versus score for (a) observer 1, (b) observer 2, (c) observer 3, and (d) all observers. Binned score for value N refers to the observer scores greater than (N – 10) and smaller than N.
|
|
|
DISCUSSION
|
|---|
A key objective in dose reduction studies is to substantiate a balance between image quality and dose. This balance is not always clear, particularly for subtle lesions. Investigation of this balance will likely cause problems because of ethical considerations, methodological issues, and rare occurrence of subtle lesions. Lesion simulation techniques can be a very valuable tool in overcoming these problems. Currently existing simulation techniques, however, have the shortcoming of the inability to produce realistic lesions. This study showed that superimposition of digital masks with proper characteristics on pediatric CT images yielded simulated lesions that experienced radiologists could not distinguish from real liver lesions.
P values consistent with expected sampling variation signified that we failed to reject the null hypothesis of a zero difference in population mean score between real and simulated lesions. This result indicated that real and simulated lesions could not be distinguished by any observer. The insignificant values of the intraclass correlation and the Kendall correlation coefficients indicated low interobserver correlation; that is, the observers did not agree on the relative scores assigned to two lesions. This was expected, given the nature of the discrimination task.
For all three observers, scores for both real and simulated lesions were similar. This was clearly shown by the small values of the mean score difference. This fact also indicated that the observers could not tell real and simulated lesions apart. If the observers could tell the lesions apart, these score differences would have been close to 100. This point was strengthened when the mean score differences for all three observers were averaged; the difference was even smaller. In fact, observers 1 and 3 assigned slightly higher scores to simulated lesions; that is, simulated lesions appeared more real than actual real lesions. All scores had large variances, and this finding indicated that observers were appropriately using the full scale of scoring in rating lesions. The fact that the differences between maximum and minimum scores for both types of lesions for all observers were very close demonstrates that the observers had no bias in assigning scores to the lesions.
All histograms also showed graphically that observers could not differentiate real from simulated lesions. If observers could have differentiated the lesions, the scores would have been separated into two peaks, a lower value peak for simulated lesions and a higher value peak for real lesions. Overall, these results affirm that, within the constraints of the experiment, simulated and real lesions were indistinguishable from each other.
Notwithstanding the preceding conclusions, there were certain limitations in our study. None of the lesions were confirmed with pathologic analysis; however, our method does mirror clinical practice, where diagnostic or treatment decisions about small liver lesions in the setting of known disease (eg, malignancy or immunodeficiency) are usually presumptive. In this study, we also used a small data set. Although no difference was shown in this data set, this finding does not necessarily suggest that there would not be any difference if we were to use a bigger data set. If there were pronounced differences between real and simulated lesions, however, those differences would be expected to be reflected in our study. Moreover, the small size of our data set is consistent with the fact that pediatric CT images containing isolated small liver lesions are not common. This point reemphasizes the importance of being able to simulate realistic small liver lesions. Other limitations include the fact that the lesions were simulated as circularly symmetric objects. Although this symmetry seemed adequate in this study, our approach can be readily extended to irregularly shaped lesions. Finally, lesions were simulated as two-dimensional objects in single sections. There is work underway to simulate lesions as three-dimensional objects in a sequence of CT images.
Within the experimental setting of this study, the results demonstrate that there was no perceptible difference between real and simulated lesions. This finding indicates that mathematic simulation of realistic lesions is a feasible technique for use in several areas of research in pediatric CT. First, in conjunction with existing simulated tube current reduction software, simulated lesions can be used for CT dose reduction studies (20). The location and presence of lesions can be controlled. With the use of simulated tube current reduction software, thresholds for small liver lesion detection can be determined. In addition, as opposed to the difficult and time-consuming process of collecting data from children with subtle lesions, large databases can be created quickly. Lesions can be inserted on images obtained in specific clinical populations, such as a particular age group or a population with a background of diffuse liver disease (eg, fatty infiltration or cirrhosis). Finally, we believe this technique can be adapted to the study of a wide range of simulated lesions in other organs.
In conclusion, we believe our simple and versatile technique for simulating realistic small liver lesions will enable the investigation of the balance between radiation dose reduction and image quality in pediatric CT to be performed in a more realistic manner than in the past.
|
Advances in Knowledge
|
|---|
- Realistic simulation of small liver lesions is feasible for pediatric CT.
- Simulation of liver lesions is a valuable tool for pediatric CT dose reduction studies.
|
ACKNOWLEDGMENTS
|
|---|
We thank George Bisset III, MD, Caroline Hollingsworth, MD, and Ana Gaca, MD, for participating in the observer study and Robert Saunders, Jr, MA, for assistance with the graphical user interface.
|
FOOTNOTES
|
|---|
Abbreviations: ROI = region of interest
See Materials and Methods for pertinent disclosures.
Author contributions: Guarantors of integrity of entire study, C.L.H., E.S., D.P.F.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; literature research, C.L.H., D.P.F.; expermimental studies, C.L.H., E.S., D.P.F.; statistical analysis, C.L.H., E.S., D.M.D.; and manuscript editing, all authors
|
References
|
|---|
- Fox SH, Toth T. Dose reduction on GE CT scanners. Pediatr Radiol 2002;32:718–723.
- Frush DP, Donnelly LF. Helical CT in children: technical considerations and body applications. Radiology 1998;209:37–48.
- Donnelly LF, Donald PF. Pediatric multidetector body CT. Radiol Clin North Am 2003;41:637–655.
- Mettler FA, Wiest PW, Locken JA, et al. CT scanning: patterns of use and dose. J Radiol Prot 2000;20:353–359.
- Linton OW, Mettler FA. National conference on dose reduction in computed tomography, emphasis on pediatrics. AJR Am J Roentgenol 2003;181:321–329.
- Hughes JS, O'Riordan MC. Radiation exposure to the UK population: 1993 review. Oxon, England: National Radiation Protection Board, 1993.
- 2005 Recommendations of ICRP. International Commission on Radiological Protection Web site. http://www.icrp.org/docs/2005_recs_CONSULTATION_Draft1a.pdf. Accessed October 23, 2005.
- Laghi A, Iannaccone R, Mangiapane F, Piacentini F, Iori S, Passariello R. Experimental colonic phantom for the evaluation of the optimal scanning technique for CT colonography using a multidetector spiral CT equipment. Eur Radiol 2003;13:459–466.
- Meaney TF, Raudkivi U, McIntyre WJ, et al. Detection of low-contrast lesions in computed body tomography: an experimental study of simulated lesions. Radiology 1980;134:149–154.
- Naidich DP, Rusinek H, McGuinness G, Leitman B, McCauley DI, Henschke CI. Variables affecting pulmonary nodule detection with computed tomography: evaluation with three-dimensional computer simulation. J Thorac Imaging 1993;8:291–299.
- Zhao B, Gamsu G, Ginsberg MS, Jiang L, Schwartz LH. Automatic detection of small lung nodules on CT utilizing a local density maximum algorithm. J Appl Clin Med Phys 2003;4:248–260.
- Boas ML. Mathematical methods in the physical sciences. New York, NY: Wiley, 1966; 233.
- Samei E, Flynn MJ, Eyler WR. Simulation of subtle lung nodules in projection chest radiography. Radiology 1997;202:117–124.
- Samei E, Flynn MJ, Eyler WR. Detection of subtle lung nodules: relative influence of quantum and anatomic noise on chest radiographs. Radiology 1999;213:727–734.
- Burgess AE, Li X, Abbey CK. Nodule detection in two component noise: toward patient structure. In: Kundel HL, ed. Proceedings of SPIE: medical imaging—imaging perception. Vol 3036. Bellingham, Wash: International Society for Optical Engineering, 1997; 2–13.
- Saunders RS Jr, Samei E, Hoeschen C. Impact of resolution and noise characteristics of digital radiographic detectors on the detectability of lung nodules. Med Phys 2004;31:1603–1613.
- Shrout PE, Fleiss JL. Intraclass correlations: uses in assessing rater reliability. Psychol Bull 1979;86:420–428.
- Fleiss JL. The design and analysis of clinical experiments. New York, NY: Wiley, 1986; 17–28.
- Conover WJ. Practical nonparametric statistics. 3rd ed. New York, NY: Wiley, 1998; 318.
- Frush DP, Slack CC, Hollingsworth CL, et al. Computer-simulated radiation dose reduction for abdominal CT of pediatric patients. AJR Am J Roentgenol 2002;179:1107–1113.
TOP
ABSTRACT
INTRODUCTION
