HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS		QUICK SEARCH:   [advanced] Author: Keyword(s): Year:  Vol:  Page:

This Article Abstract Full Text Full Text (PDF) Appendix Submit a response Alert me when this article is cited Alert me when eLetters are posted Alert me if a correction is posted Citation Map Services Email this article to a friend Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Rubin, G. D. Articles by Napel, S. Search for Related Content PubMed PubMed Citation Articles by Rubin, G. D. Articles by Napel, S.

Pulmonary Nodules on Multi–Detector Row CT Scans: Performance Comparison of Radiologists and Computer-aided Detection¹

¹ From the Departments of Radiology (G.D.R., J.K.L., D.S.P., L.C.C., A.N.L., R.M., P.K.S.D., S.E.Z., S.N.) and Electrical Engineering (A.J.S.), Stanford University School of Medicine, 300 Pasteur Dr, S-072, Stanford, CA 94305-5105; and Department of Radiology, New York University School of Medicine, New York, NY (D.P.N.). Received March 31, 2004; revision requested June 8; revision received July 26; accepted August 19. Address correspondence to G.D.R. (e-mail: grubin@stanford.edu).

View larger version (71K): [in a new window]   Figure 1. Application of surface normal overlap (SNO)-CAD and lantern transform to nodule detection in human lung. A, Detail of transverse CT section shows pulmonary nodule in posterior portion of right upper lobe. Applied threshold indicates isosurface (red) between air and tissue. B, Schematic representation of surface normal vectors (blue) generated at isosurface (red) in same lung region as that bounded by aqua box in A. Normals are generated in three dimensions in volumetric CT data at all isosurfaces in the thick region extracted during segmentation. C, Map of SNO-CAD scores for all voxels displayed in A, calculated by using a clustering algorithm to quantify convergence of normal vectors, confirms identification of pulmonary nodule. Scale at bottom indicates 12-bit scaling of SNO-CAD scores (0-8). D, Image with superimposed schema shows application of lantern transform to nodule in contact with blood vessel. Rays of visibility (black radii) cast from a SNO-CAD-identified nodule candidate are used to generate an approximately spherical surface (red). E, Image with superimposed schema shows application of lantern transform to SNO-CAD-identified nodule candidate in pulmonary vessel. Rays of visibility generate an ellipsoid surface less spherical than that in D. Quantitative characteristics of this ellipsoid result in its rejection as a nodule candidate, while those of the ellipsoid in D allow it to remain a nodule candidate.

View larger version (17K): [in a new window]   Figure 2. Bar graph shows the per-patient distribution of 195 noncalcified nodules with a diameter of 3 mm or more.

View larger version (23K): [in a new window]   Figure 3. Bar graph shows the distribution of 195 noncalcified nodules with a diameter of 3 mm or more, according to diameter range.

View larger version (34K): [in a new window]   Figure 4. Pie chart shows the distribution of 936 FP detections, determined by the consensus panel to be definitely not nodules, according to the nature of the finding.

View larger version (28K): [in a new window]   Figure 5. Free-response ROC plots show sensitivity versus FP detections per patient (log scale) for radiologists and the cross-validation-trained CAD system among, A, noncalcified nodules with a diameter of 3 mm or more and, B, noncalcified nodules with a diameter of 5 mm or more. The legend in B applies also to A, but note the difference between A and B in the scaling of the y-axis.

View larger version (28K): [in a new window]   Figure 6. Alternative free-response ROC curves for each reader and CAD. Areas under the curves (A1) are given. P(FPI) = probability of a FP identification.

View larger version (42K): [in a new window]   Figure 7. Venn diagrams demonstrate variability among the three readers in the number of TP detections with a confidence level of 3-5 for, A, nodules with a diameter of 3 mm or more and, B, nodules with a diameter of 5 mm or more. Inside the circles, numbers in parentheses represent TP detections that were not detected by the CAD system. Inside the boxes, numbers indicate total TP detections with CAD used at an upper threshold of 15 FP detections per patient; numbers in parentheses indicate TP detections by the CAD system that were not seen by any reader. These diagrams show that at the selected thresholds all three readers made unique detections and that only 33% of nodules with a diameter of 3 mm or more and 64% of nodules with a diameter of 5 mm or more were detected by all three readers. The CAD system detected a majority of the readers’ detections and made 35 unique detections (3-mm-diameter nodules)—more than any reader.

View larger version (25K): [in a new window]   Figure 8. Bar graph shows mean sensitivities for the three readers individually, paired in double readings, and paired with the CAD system, assuming 100% acceptance of TP CAD detections by the radiologists. Four CAD thresholds were tested and are indexed according to the mean number of FP detections that readers would need to assess and exclude. Only nodules identified with a reader confidence level of 3-5 were included in this analysis. The error bars indicate the low-high range of the values. This diagram illustrates that when CAD operates at a threshold that results in only three FP detections per patient, significantly better performance and reduced interradiologist variability (narrower high-low error bars) were found for radiologists paired with the CAD system, compared with radiologists reading alone or in pairs. Note the small incremental improvement in performance as CAD is allowed to become more sensitive at the expense of a greater number of FP detections.