DOI: 10.1148/radiol.2423060260
(Radiology 2007;242:811-816.)
© RSNA, 2007
Lytic Metastases in Thoracolumbar Spine: Computer-aided Detection at CT—Preliminary Study1
Stacy D. O'Connor, SB,
Jianhua Yao, PhD and
Ronald M. Summers, MD, PhD
1 From the Diagnostic Radiology Department, Clinical Center, National Institutes of Health (NIH), Bldg 10, Room 1C351, 10 Center Dr, MSC 1182, Bethesda, MD 20892-1182. Received February 17, 2006; revision requested April 21; revision received June 14; final version accepted July 7. Supported in part by the Intramural Research Program of the NIH, Clinical Center. Made possible through the Clinical Research Training Program, a public-private partnership supported jointly by the NIH and a grant to the Foundation for the NIH from Pfizer Pharmaceuticals Group.
Address correspondence to R.M.S. (e-mail: rms{at}nih.gov).
 |
ABSTRACT
|
|---|
Purpose: To evaluate the sensitivity of a computer-aided detection (CAD) system for detection of lytic thoracolumbar spinal lesions at body CT, with results of manual lesion segmentation as the reference standard.
Materials and Methods: The study was HIPAA compliant and institutional review board approved; the institutional review board waived the need for informed consent. The CAD system segments the spine on CT images and searches for detections that match size, shape, location, and attenuation criteria. To reduce false-positive findings, 16 features for each detection were computed and fed to a classifier trained with manually segmented lesions. The data set consisted of CT studies of 50 patients (30 men, 20 women; range, 18–82 years; mean, 54.8 years) with 28 lesions. Studies were assigned to either a training (29 studies) or testing (21 studies) set. Sensitivities and false-positive rates (FPRs) for training and testing sets were calculated for these lesions, which were probable lytic metastases with areas 0.8 cm2 or greater.
Results: Training set sensitivity was 0.83 (10 of 12; 95% confidence interval: 0.51, 0.97), with an FPR of 7.4 per patient. Test set sensitivity was 0.94 (15 of 16; 95% confidence interval: 0.68, 1.00), with an FPR of 4.5 per patient. There was no significant difference between the CAD sensitivities of the training and test sets (P = .56). Of three false-negative findings, two were due to incomplete segmentation of the vertebral pedicle, and the third was rejected by the classifier. False-positive detections were most often attributable to veins that connect the basivertebral vein with the anterior venous plexus (106 [34%] of 310) and to low-attenuating disks (83 [27%] of 310).
Conclusion: This CAD system successfully identified probable lytic metastases in the thoracolumbar spine and generalized well to an independent testing set.
Supplemental material: http://radiology.rsnajnls.org/cgi/content/full/242/3/811/DC1
© RSNA, 2007
|
INTRODUCTION
|
|---|
Bone metastases can appear lytic, sclerotic, or anywhere in a continuum between these extremes. Therefore, the automated detection of these lesions is a complex problem that must be broken down into manageable components. Lytic lesions are more likely to cause pathologic fractures and higher baseline serum calcium levels than are mixed or sclerotic lesions (1,2). Early diagnosis of these spinal metastases is important because treatment before the development of substantial morbidities improves outcomes (3,4). For example, one breast cancer study (5) revealed that nearly all patients were likely to have manifestation of bone metastases prior to development of spinal cord compression. Ninety-six percent of patients who were diagnosed early, with the ability to walk, retained this ability after treatment, whereas only 45% of those who were diagnosed late in the disease, when they were unable to walk, regained the ability after treatment. Because of their deleterious nature, we focused on lytic lesions in this study.
Although there are few data on the subject, it has been postulated that bone metastases are missed at CT (6). Findings of bone metastases can be subtle and must be recognized even when a radiologist has not been directed to look for them. To do this, radiologists should use bone windows, but there is a perception that these windows are underused by radiologists and oncologists (6).
Computer-aided detection (CAD) is a software tool that can indicate potential abnormalities for further scrutiny by a radiologist. It has been used to detect lesions in the breast, where it increased the true-positive rate in breast cancer screening and improved the yield of biopsy recommendations for patients with masses on serial mammograms (7,8). CAD has been shown to improve radiologist performance in detecting lung nodules at chest radiography and CT (9,10) and to increase sensitivity for detecting polyps at CT colonography (11). To our knowledge, CAD for bone metastases has not yet been evaluated. The development of CAD of bone metastases would enhance the future global or all-purpose CAD toolbox envisioned in recent editorials (12,13). Thus, the purpose of our study was to evaluate the sensitivity of a CAD system for detection of lytic thoracolumbar spinal lesions at body CT, with results of manual lesion segmentation as the reference standard.
|
MATERIALS AND METHODS
|
|---|
Subjects
The study was Health Insurance Portability and Accountability Act compliant and institutional review board approved. The institutional review board waived the need for informed consent. The database of radiology reports from our institution was searched with the following parameters: age, 18–100 years; date of examination, January 1 to December 31, 2004; and the keywords: lytic, lucent, lucency, or lucencies, and CT chest, abdomen, and/or pelvis. One author (S.D.O.) reviewed the results of this search and selected patients with at least one lytic lesion in the thoracic or lumbar vertebrae (Fig 1). Patients with extensive disease (eg, obliterated vertebrae or more than 20 lesions) or hardware such as cement, rods, or screws were excluded from the study. If patients had undergone multiple CT examinations during our search interval, we used the first CT examination encountered. To maintain consistency, only intravenous contrast material–enhanced studies were used.
View larger version (12K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 1: Patient inclusion and exclusion criteria flowchart. RIS = radiology information system, TLS = thoracolumbar spinal. * = Other exclusion criteria, which included not being first CT examination; presence of hardware, extensive disease, or extensive scoliosis; no intravenous contrast material; image problems (missing images, change in section thickness or spacing within scan); or different scanner used. (One examination performed with a Philips scanner was included because it was initially misidentified.)
|
|
The resultant study group consisted of 50 patients (30 men, 20 women; age range, 18–82 years; mean age, 54.8 years) whose studies were divided into training and test cases (29 and 21 patients, respectively). We first collected all the training cases in the CAD development phase and then collected all the test cases in the CAD testing phase. The test cases dated from January to June 2004, and the training cases dated from July to December 2004. Patients had been given a diagnosis of melanoma (n = 19), renal cell carcinoma (n = 10), prostate cancer (n = 4), lung cancer (n = 4), lymphoma (n = 2), breast cancer (n = 2), pheochromocytoma (n = 2), or other disorders (n = 7).
CT Scanning
Each patient was scanned with either a four-detector (LightSpeed QX/I; GE Healthcare, Waukesha, Wis) (20 patients), eight-detector (LightSpeed Ultra; GE Healthcare) (29 patients), or 16-detector (Mx8000 IDT; Philips, Andover, Mass) (one patient) CT scanner. Images were obtained with a 5-mm section thickness. Data sets consisted of an average of 124 images (range, 60–145 images). Patients underwent CT of the chest, abdomen, and pelvis (44 patients); abdomen and pelvis (two patients); chest (one patient); chest and abdomen (one patient); abdomen (one patient); or pelvis (one patient). The standard reconstruction kernel was used for 49 patients, and the B kernel was used for one patient. The patients received 110–130 mL iopamidol (Isovue 300; Bracco Diagnostics, Princeton, NJ) injected intravenously with a power injector (EnVision; Medrad, Indianola, Pa).
Identifying Lesions
CT images were transferred to a personal computer (Precision 530; Dell, Austin, Tex) and reviewed with Medical Image Processing, Analysis, and Visualization software (14) for lesions 5 mm or larger. By using a two-reader consensus (S.D.O., a trained 4th-year medical student and R.M.S., a board-certified radiologist with 11 years of experience reading body CT images), all thoracolumbar spinal lesions were identified and qualitatively characterized as lytic, sclerotic, or mixed (at least 20% lytic voxels in a predominately sclerotic lesion, at least 20% sclerotic voxels in a predominately lytic lesion, or lesion without a predominant voxel type). Lytic lesions were further characterized as probable or unlikely metastases; only probable metastases were included in our study. For example, unlikely metastases included Schmorl nodes, areas of degenerative disk disease or osteopenia, and hemangiomas.
The largest lytic area of each lesion was measured by using the largest dimension in the x-y plane; if lesions merged at some point, they were considered one lesion. Because of the 5-mm-thick sections, the z plane was not used. Lytic area was calculated by using all voxels within the lesion with an attenuation of less than 176 HU, which was an arbitrary cutoff value defined on the basis of visual observation. To select lesions of substantial size, we set a minimum lytic-area threshold of 0.8 cm2, which corresponds to a circle with a diameter of more than 1 cm. There were 28 probable lytic metastases (12 in the training set, 16 in the test set) with lytic areas greater than the threshold value. Patients had between zero and four probable lytic metastases (average, 0.6) with areas ranging from 0.9 to 10.6 cm2 (average, 2.7 cm2). Thirty-three patients with lytic lesions did not have any lesions characterized as probable metastases. Seven patients had more than one probable lytic metastasis (two in the training set, five in the test set). A total of 35 mixed lesions (24 in the training set, 11 in the test set) and 37 sclerotic lesions (23 in the training set, 14 in the test set) were also present in our study group.
Lesions were manually segmented by a trained student (S.D.O.), supervised by a board-certified radiologist (R.M.S.), who drew a contour along the voxels on the edge of each lesion on each section on which it appeared. Results of manual segmentation were used as the reference standard segmentation in our study.
Computer-aided Lesion Detection Method
The computer-aided bone lesion detection method has three stages: segmentation, feature computation, and classification. More details have been published elsewhere (15). See Appendix E1 (http://radiology.rsnajnls.org/cgi/content/full/242/3/811/DC1).
The training set was used to train the classifier. The classifier then was applied to test data to assess the classifier's ability to distinguish true-positive findings from false-positive findings.
Assessment of the Method
The CAD system was validated by matching the three-dimensional computer detections with manually segmented reference-standard lesions. If a CAD detection had any voxels in common with a manual detection, it was considered a true detection; otherwise, it was considered a false detection. If two separate CAD detections overlapped a single reference-standard detection, we counted one true-positive finding; the CAD method was not "penalized" by counting one of the detections as a false-positive finding. Because we were not attempting to detect mixed or sclerotic lesions, we considered only detections that matched probable lytic metastasis reference-standard detections and removed those that matched mixed and sclerotic reference-standard detections from the analysis. Our reference-standard segmentation included only lesions in the thoracolumbar spine, so we identified the sections on which the thoracic spine began and on which the lumbar spine ended and removed from consideration detections that did not fall between these sections.
Statistical Analysis
Sensitivity (the fraction of detections that overlapped with reference standard detections) and false-positive rate (FPR) per study were used as the overall means of assessment of our method. Sensitivity and FPR were reported for probable lytic metastases larger than 0.8 cm2 (equivalent to circular lesions of 1 cm in diameter) for both the training and test sets. The causes of false-positive and false-negative findings were assessed. Confidence intervals for sensitivity were computed with an online calculator (16). The Fisher exact test was used to compare the sensitivities between the test and training sets (17). Free-response receiver operating characteristic curves were calculated for the CAD system for the training and test sets. Sensitivities and FPRs were reported at a single operating point on the free-response receiver operating characteristic curve. The operating point is determined in the CAD software by setting a threshold on a numeric value output by the classifier for each detection. P 
.05 was considered to indicate a statistically significant difference.
|
RESULTS
|
|---|
The training set had a sensitivity of 0.83 (10 of 12; 95% confidence interval: 0.51, 0.97) per lesion and an FPR of 7.4 ± 5.3 (standard deviation) per patient for probable lytic metastases. Sensitivity of the test set was 0.94 (15 of 16; 95% confidence interval: 0.68, 1.00) per lesion, with an FPR of 4.5 ± 1.9 per patient. The difference in sensitivity between the training and test sets was not statistically significant (P = .56) (Figs 2–4).
View larger version (97K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2a: CT images of true-positive CAD of 1.9-cm-diameter (area, 1.9-cm2) lytic lesion in L1 vertebral body in 58-year-old woman with melanoma. (a) Original transverse bone window (window width, 2500 HU; window level, 480 HU) image shows lesion (arrow). (b) Color-encoded transverse image (green = CAD marks, red = spine recognition with CAD software, blue = spinal canal recognition, brown = lytic lesion candidate rejected by filter or classifier) shows true-positive CAD mark and shows false-positive finding in osteopenic area (large arrow). Also indicated are true-negative findings (small arrows), which were detected but eliminated by the classifier.
|
|
View larger version (85K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2b: CT images of true-positive CAD of 1.9-cm-diameter (area, 1.9-cm2) lytic lesion in L1 vertebral body in 58-year-old woman with melanoma. (a) Original transverse bone window (window width, 2500 HU; window level, 480 HU) image shows lesion (arrow). (b) Color-encoded transverse image (green = CAD marks, red = spine recognition with CAD software, blue = spinal canal recognition, brown = lytic lesion candidate rejected by filter or classifier) shows true-positive CAD mark and shows false-positive finding in osteopenic area (large arrow). Also indicated are true-negative findings (small arrows), which were detected but eliminated by the classifier.
|
|
View larger version (105K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 3a: CT images of false-positive CAD system detections. (a, b) Detection on basivertebral vein (arrow) in L4 vertebral body in 74-year-old man with melanoma. (c, d) Detection in T12-L1 intervertebral disk (arrow). (a, c) Original transverse CT images (window width, 2500 HU; window level, 480 HU). (b, d) Color-encoded images (green = CAD marks, red = spine recognition with CAD software, blue = spinal canal recognition, brown = lytic lesion candidate rejected by filter or classifier).
|
|
View larger version (82K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 3b: CT images of false-positive CAD system detections. (a, b) Detection on basivertebral vein (arrow) in L4 vertebral body in 74-year-old man with melanoma. (c, d) Detection in T12-L1 intervertebral disk (arrow). (a, c) Original transverse CT images (window width, 2500 HU; window level, 480 HU). (b, d) Color-encoded images (green = CAD marks, red = spine recognition with CAD software, blue = spinal canal recognition, brown = lytic lesion candidate rejected by filter or classifier).
|
|
View larger version (85K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 3c: CT images of false-positive CAD system detections. (a, b) Detection on basivertebral vein (arrow) in L4 vertebral body in 74-year-old man with melanoma. (c, d) Detection in T12-L1 intervertebral disk (arrow). (a, c) Original transverse CT images (window width, 2500 HU; window level, 480 HU). (b, d) Color-encoded images (green = CAD marks, red = spine recognition with CAD software, blue = spinal canal recognition, brown = lytic lesion candidate rejected by filter or classifier).
|
|
View larger version (74K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 3d: CT images of false-positive CAD system detections. (a, b) Detection on basivertebral vein (arrow) in L4 vertebral body in 74-year-old man with melanoma. (c, d) Detection in T12-L1 intervertebral disk (arrow). (a, c) Original transverse CT images (window width, 2500 HU; window level, 480 HU). (b, d) Color-encoded images (green = CAD marks, red = spine recognition with CAD software, blue = spinal canal recognition, brown = lytic lesion candidate rejected by filter or classifier).
|
|
View larger version (13K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 4: Graph of free-response receiver operating characteristic curves for CAD system for detection of probable lytic metastatic bone lesions at least 0.8 cm2 in cross-sectional area. Sensitivity per lesion and FPRs per patient are shown.  = Operating points. CAD system performed similarly on training and test sets.
|
|
The reasons for the 310 false-positive detections in the training and test sets could be broken down into the following seven main categories: (a) peripheral vein—venous connections between the basivertebral vein and the anterior external venous plexus (n = 106 [34%]); (b) disk—low-attenuating disks or volume averaging with disks (n = 83 [27%]); (c) osteopenia (n = 68 [22%]); (d) outside—detection on areas outside the vertebra (n = 37 [12%]); (e) basivertebral vein—a basivertebral vein that enters the posterior vertebral body (n = 6 [2%]); (f) normal—a decrease in attenuation from volume averaging with normal structures such as joints or from oblique cuts through the cortex (n = 6 [2%]); and (g) spinal canal—software mistook part of the spinal canal for lytic vertebral body lesions due to artifacts on images (n = 2 [1%]). Two of the false-positive detections were actually of reference-standard lesions that were not segmented on all sections in which they appeared. False-positive detections varied greatly among patients, numbering 0–20 per patient (average, 6.2).
We also analyzed the three false-negative detections (two in the training set, one in the test set). Two were in pedicles that were not properly segmented, so the lesions were never detected. The other false-negative finding was initially detected but thrown out by the classifier, most likely due to similarity to a basivertebral vein.
|
DISCUSSION
|
|---|
We have described a CAD system that successfully detects lytic lesions in the thoracolumbar spine at CT. Sensitivity and FPR were 0.94 and 4.5, respectively, for probable lytic metastases in the test set.
Although CT is not the modality of choice for whole-body screening for bone metastases—skeletal scintigraphy or conventional radiography are the most common choices (6,18–21)—metastases still must be identified at CT when possible. When bone window settings are used, CT shows a high level of detail in bone, enabling one to distinguish among materials of different attenuations (22,23). For depicting metastases to the spine, CT is superior to skeletal scintigraphy and conventional radiography (24,25) and has reported sensitivities ranging from 93% to 100% (26–28). However, these lesions can be subtle and easily overlooked by a radiologist, especially when bone windows are underused (6) and the radiologist has not been specifically directed by the referring physician to look for metastases.
Detecting spinal lesions with a computer is challenging, owing to the variation in bone attenuation within and among patients, as well as to the diversity of nonmetastatic abnormalities such as degenerative disk disease. The problem must be broken down into manageable components that can be addressed sequentially. Because of the deleterious nature of lytic spinal metastases, we decided to focus on that subclass in the initial development of our CAD system. Because this system was not intended to depict mixed or sclerotic lesions, we do not present results regarding these other lesion subclasses.
The first task in detecting lytic spinal metastases is to locate and segment the spine and exclude other structures. This is most difficult in the thoracic spine, where the ribs are often depicted along with the vertebrae. We included location criteria in our filter and classifier to account for this, but a number of the outside false-positive detections were in costovertebral joints. These are low-attenuating regions surrounded by high-attenuating cortex. The synovial joints between adjacent vertebrae and the nearby contrast material–filled inferior vena cava are sometimes segmented along with the lumbar vertebrae, which results in outside non–bone-related false-positive findings.
While adjacent high-attenuating structures pose a challenge for segmentation of the spine from nonspinal structures, the intrinsically low-attenuating intervertebral disks pose a challenge during both segmentation and lesion detection. Disks and lytic lesions are both low attenuating, so disks may be mistaken for lesions and may cause volume averaging with the vertebrae. This accounted for 27% of false-positive detections in the training set. In the future, these false-positive detections may be reduced if the intervertebral disks can be automatically identified within the spinal column.
The 5-mm section thickness of our data set posed another challenge to the segmentation of the spine and detection of lesions. This thickness is common for routine CT of the chest, abdomen, and/or pelvis. Our system is designed to find unexpected lytic spinal metastases at examinations ordered for other indications. The use of thick sections leads to volume averaging, causing parts of the vertebral body to have attenuation similar to that of surrounding soft tissues. Substantial leakage (segmentation of undesired structures) can happen when region-based segmentation is applied. We adopted a multipass technique to address this problem. First, a high threshold was applied to achieve the initial segmentation, then morphologic operations and "rolling balls" were applied to close the holes and gaps. The use of thick sections increases normal false-positive detections, especially in the vertebral arch, when oblique cuts result in volume averaging of vertebral cortex and adjacent soft tissue. Finally, the section thickness makes lesions more difficult to detect, because most appear on only one section.
Other low-attenuating structures that cause false-positive detections are the spinal canal, the basivertebral vein, and the basivertebral vein's connections to the anterior external venous plexus. These false-positive findings have characteristic features, especially in terms of location and shape, which may be used in future CAD systems to recognize and eliminate them from the CAD potential lesion list that is presented to radiologists for consideration.
Our false-negative results can be attributed to two main causes. Two were due to failure in segmentation of the pedicle, while the final false-negative finding was due to failure in characterization. Further work could increase the accuracy of segmentation of the pedicle, while a larger training set may help the classifier distinguish between true-positive and false-positive lesions.
This study had several limitations. First, we did not have biopsy or follow-up information for the patients and therefore could not confirm the pathologic findings of the lesions. This is not unexpected, because metastatic lesions are typically not biopsed unless biopsy results are required for local therapy or initial diagnosis. Many patients did not undergo follow-up studies within the course of our study, even when suspicious lesions were identified. Therefore, we used a two-reader consensus for the probable pathologic findings of a lesion. Because CT can be used to characterize the cause of spinal lesions, it is not unreasonable to rely on this examination (25,29). Second, our patient pool was biased toward patients in whom radiologists noted lytic areas in the spinal column during routine workup. This may mean our system was trained and tested on lesions that were more obvious than the average lytic metastasis. However, we reanalyzed all scans in this study and used all probable metastases in our analysis, including the more subtle lesions.
Third, lesion identification was performed by consensus of one radiologist and one 4th-year medical student. Adding additional radiologists might have improved the reference standard. Fourth, subtle lesions might have been missed, and lesion conspicuity was not assessed. Fifth, matching reference-standard and CAD system detections were recorded as true-positive findings if there was any overlap whatsoever between the two detections. In future CAD experiments, a greater amount of overlap may be required for a true-positive finding. Sixth, seven patients had more than one probable lytic metastasis, and this clustering of lesions was not addressed. Future investigators of larger patient populations may be able to select patients with only one probable lytic metastasis to avoid clustering. Finally, we analyzed performance in the subset of lesions with lytic areas greater than 0.8 cm2. Because this was an initial study, large lesions were an appropriate target; future work should include smaller lesions.
In conclusion, we have described a CAD system that can detect probable lytic metastases in the thoracolumbar spine. We have identified some of the common causes of false-negative and false-positive detections to guide further development of bone CAD systems. Future bone CAD systems are likely to detect lesions in other bones and to find mixed and sclerotic lesions. Additional research will be required to show whether bone CAD systems improve radiologist diagnostic accuracy and interpretive efficiency.
|
ADVANCES IN KNOWLEDGE
|
|---|
- A computer-aided detection (CAD) algorithm for lytic bone metastases in the thoracolumbar spine at body CT has a high sensitivity at a clinically reasonable false-positive rate.
- Performance of the CAD system was similar on the training and test sets, indicating that the CAD system generalizes well to new data.
- We identified common causes of false-positive findings (veins and intervertebral disks) that can guide further development of the CAD system.
|
ACKNOWLEDGMENTS
|
|---|
We thank Andrew Dwyer, MD, for critical review of the paper.
|
FOOTNOTES
|
|---|
Abbreviations: CAD = computer-aided detection • FPR = false-positive rate
Authors stated no financial relationship to disclose.
Author contributions: Guarantors of integrity of entire study, all authors; 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, S.D.O., J.Y.; clinical studies, all authors; statistical analysis, all authors; and manuscript editing, all authors
|
References
|
|---|
- Rubens RD. Bone metastases: the clinical problem. Eur J Cancer 1998;34:210–213.[CrossRef][Medline]
- Mercadante S. Malignant bone pain: pathophysiology and treatment. Pain 1997;69:1–18.[CrossRef][Medline]
- Bilsky MH, Lis E, Raizer J, Lee H, Boland P. The diagnosis and treatment of metastatic spinal tumor. Oncologist 1999;4:459–469.[Abstract/Free Full Text]
- Colman LK, Porter BA, Redmond J 3rd, et al. Early diagnosis of spinal metastases by CT and MR studies. J Comput Assist Tomogr 1988;12:423–426.[Medline]
- Hill ME, Richards MA, Gregory WM, Smith P, Rubens RD. Spinal cord compression in breast cancer: a review of 70 cases. Br J Cancer 1993;68:969–973.[Medline]
- Hamaoka T, Madewell JE, Podoloff DA, Hortobagyi GN, Ueno NT. Bone imaging in metastatic breast cancer. J Clin Oncol 2004;22:2942–2953.[Abstract/Free Full Text]
- Irwig L, Houssami N, van Vliet C. New technologies in screening for breast cancer: a systematic review of their accuracy. Br J Cancer 2004;90:2118–2122.[Medline]
- Hadjiiski L, Chan HP, Sahiner B, et al. Improvement in radiologists' characterization of malignant and benign breast masses on serial mammograms with computer-aided diagnosis: an ROC study. Radiology 2004;233:255–265.[Abstract/Free Full Text]
- Awai K, Murao K, Ozawa A, et al. Pulmonary nodules at chest CT: effect of computer-aided diagnosis on radiologists' detection performance. Radiology 2004;230:347–352.[Abstract/Free Full Text]
- Kakeda S, Moriya J, Sato H, et al. Improved detection of lung nodules on chest radiographs using a commercial computer-aided diagnosis system. AJR Am J Roentgenol 2004;182:505–510.[Abstract/Free Full Text]
- Summers RM, Jerebko AK, Franaszek M, Malley JD, Johnson CD. Colonic polyps: complementary role of computer-aided detection in CT colonography. Radiology 2002;225:391–399.[Abstract/Free Full Text]
- Summers RM. Road maps for advancement of radiologic computer-aided detection in the 21st century. Radiology 2003;229:11–13.[Free Full Text]
- Krupinski EA. Computer-aided detection in clinical environment: benefits and challenges for radiologists. Radiology 2004;231:7–9.[Free Full Text]
- McAuliffe MJ, Lalonde FM, McGarry D, Gandler W, Csaky K, Trus BL. Medical image processing, analysis & visualization in clinical research. In: IEEE computer-based medical systems (CBMS). Bethesda, Md: Institute of Electrical and Electronic Engineers, 2001; 381–386.
- Yao J, O'Connor SD, Summers RM. Computer aided lytic bone metastasis detection using regular CT images. In: Reinhardt JM, Pluim JP, eds. Proceedings of SPIE: medical imaging 2006—Image Processing. Vol 6144. Bellingham, Wash: International Society for Optical Engineering, 2006; 1692–1700.
- Lowry R. Clinical calculator 1: from an observed sample—estimates of population prevalence, sensitivity, specificity, predictive values, and likelihood ratios. VassarStats. http://faculty.vassar.edu/lowry/clin1.html. Accessed February 7, 2006.
- Lowry R. Fisher's exact probability test. http://faculty.vassar.edu/lowry/fisher.html. Accessed February 9, 2006.
- Soderlund V. Radiological diagnosis of skeletal metastases. Eur Radiol 1996;6:587–595.[Medline]
- Rybak LD, Rosenthal DI. Radiological imaging for the diagnosis of bone metastases. Q J Nucl Med 2001;45:53–64.[Medline]
- Rosenthal DI. Radiologic diagnosis of bone metastases. Cancer 1997;80:1595–1607.[CrossRef][Medline]
- Kagan AR, Bassett LW, Steckel RJ, Gold RH. Radiologic contributions to cancer management: bone metastases. AJR Am J Roentgenol 1986;147:305–312.[Free Full Text]
- Coleman RE, Rubens RD. Bone metastases and breast cancer. Cancer Treat Rev 1985;12:251–270.[CrossRef][Medline]
- Tryciecky EW, Gottschalk A, Ludema K. Oncologic imaging: interactions of nuclear medicine with CT and MRI using the bone scan as a model. Semin Nucl Med 1997;27:142–151.[CrossRef][Medline]
- Coleman RE. Monitoring of bone metastases. Eur J Cancer 1998;34:252–259.[CrossRef][Medline]
- Sundaram M, McGuire MH. Computed tomography or magnetic resonance for evaluating the solitary tumor or tumor-like lesion of bone? Skeletal Radiol 1988;17:393–401.[CrossRef][Medline]
- Muindi J, Coombes RC, Golding S, Powles TJ, Khan O, Husband J. The role of computed tomography in the detection of bone metastases in breast cancer patients. Br J Radiol 1983;56:233–236.[Abstract]
- Daffner RH, Lupetin AR, Dash N, Deeb ZL, Sefczek RJ, Schapiro RL. MRI in the detection of malignant infiltration of bone marrow. AJR Am J Roentgenol 1986;146:353–358.[Abstract/Free Full Text]
- Metser U, Lerman H, Blank A, Lievshitz G, Bokstein F, Even-Sapir E. Malignant involvement of the spine: assessment by 18F-FDG PET/CT. J Nucl Med 2004;45:279–284.[Abstract/Free Full Text]
- Redmond J 3rd, Spring DB, Munderloh SH, George CB, Mansour RP, Volk SA. Spinal computed tomography scanning in the evaluation of metastatic disease. Cancer 1984;54:253–258.[CrossRef][Medline]
Related Article
-
Erratum
Radiology 2007 244: 320.
[Full Text]
[PDF]
TOP
ABSTRACT
INTRODUCTION
