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) 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 Chan, H.-P. Articles by Sanjay-Gopal, S. Search for Related Content PubMed PubMed Citation Articles by Chan, H.-P. Articles by Sanjay-Gopal, S.

Improvement of Radiologists' Characterization of Mammographic Masses by Using Computer-aided Diagnosis: An ROC Study¹

¹ From the Department of Radiology, University of Michigan Hospital, UH B1F510, 1500 E Medical Center Dr, Ann Arbor, MI 48109-0030. From the 1997 RSNA scientific assembly. Received August 10, 1998; revision requested September 8; revision received November 30; accepted January 21, 1999. Supported in part by United States Public Health Service grant CA 48129 and by U.S. Army Medical Research and Materiel Command grant DAMD 17-96-1-6254. B.S. supported by Career Development award DAMD 17-96-1-6012 from the U.S. Army Medical Research and Materiel Command. N.P. supported by a grant from the Whitaker Foundation. Address reprint requests to H.P.C. (e-mail: chanhp@umich.edu).

View larger version (21K): [in a new window]   Figure 1a. Histograms illustrate the distributions of (a) size (ie, length of the long axis) and (b) visibility ranking (1 = obvious, 5 = subtle) of the 253 masses included in the data set. Because classification accuracy depends on the case mix, these distributions provided some information on the masses in the data set.

View larger version (18K): [in a new window]   Figure 1b. Histograms illustrate the distributions of (a) size (ie, length of the long axis) and (b) visibility ranking (1 = obvious, 5 = subtle) of the 253 masses included in the data set. Because classification accuracy depends on the case mix, these distributions provided some information on the masses in the data set.

View larger version (168K): [in a new window]   Figure 2a. Example of rubber-band-straightening transform for extraction of texture features in the margin region surrounding a mass. (a) Original and (b) background-corrected images showing the region of interest with the mass, (c) mammogram showing an outline of the segmented mass, and (d) rubber-band-straightening-transformed image of a 40-pixel-wide region surrounding the segmented mass.

View larger version (163K): [in a new window]   Figure 2b. Example of rubber-band-straightening transform for extraction of texture features in the margin region surrounding a mass. (a) Original and (b) background-corrected images showing the region of interest with the mass, (c) mammogram showing an outline of the segmented mass, and (d) rubber-band-straightening-transformed image of a 40-pixel-wide region surrounding the segmented mass.

View larger version (178K): [in a new window]   Figure 2c. Example of rubber-band-straightening transform for extraction of texture features in the margin region surrounding a mass. (a) Original and (b) background-corrected images showing the region of interest with the mass, (c) mammogram showing an outline of the segmented mass, and (d) rubber-band-straightening-transformed image of a 40-pixel-wide region surrounding the segmented mass.

View larger version (18K): [in a new window]   Figure 2d. Example of rubber-band-straightening transform for extraction of texture features in the margin region surrounding a mass. (a) Original and (b) background-corrected images showing the region of interest with the mass, (c) mammogram showing an outline of the segmented mass, and (d) rubber-band-straightening-transformed image of a 40-pixel-wide region surrounding the segmented mass.

View larger version (25K): [in a new window]   Figure 3. Histogram of the test discriminant scores of the 253 masses obtained from the linear discriminant classifier by using a "leave one case out" training and test resampling scheme. For this classifier, a smaller discriminant score corresponded to a higher likelihood of malignancy. The discriminant scores were used as the decision variable in the ROC analysis of classification performance.

View larger version (20K): [in a new window]   Figure 4. Binormal distribution fitted to the histogram of the discriminant scores of the malignant and benign masses. The discriminant scores were linearly transformed into a relative malignancy rating ranging from 1 to 10, where 1 corresponded to the most benign rating and 10 corresponded to the most malignant rating. This binormal distribution was shown to the observers during the training session to explain the rating scale of the computer classifier.

View larger version (20K): [in a new window]   Figure 5. ROC curve for computerized classification of the 238 masses used in the observer performance study with single-view reading. The computer's ROC curve can be compared with the radiologists' ROC curves obtained from the single-view reading experiment illustrated in Figures 6 and 8.

View larger version (30K): [in a new window]   Figure 6a. ROC curves for the six observers for single-view reading of the masses (a) without CAD and (b) with CAD. (a, b) R1 = reader 1, R2 = reader 2, R3 = reader 3, R4 = reader 4, R5 = reader 5, R6 = reader 6. Five of the six observers achieved an increase in the area under the ROC curve, Az, with CAD.

View larger version (27K): [in a new window]   Figure 6b. ROC curves for the six observers for single-view reading of the masses (a) without CAD and (b) with CAD. (a, b) R1 = reader 1, R2 = reader 2, R3 = reader 3, R4 = reader 4, R5 = reader 5, R6 = reader 6. Five of the six observers achieved an increase in the area under the ROC curve, Az, with CAD.

View larger version (28K): [in a new window]   Figure 7a. ROC curves for the six observers for two-view reading of the masses (a) without CAD and (b) with CAD. (a, b) R1 = reader 1, R2 = reader 2, R3 = reader 3, R4 = reader 4, R5 = reader 5, R6 = reader 6. All six observers achieved an increase in the area under the ROC curve, Az, with CAD.

View larger version (25K): [in a new window]   Figure 7b. ROC curves for the six observers for two-view reading of the masses (a) without CAD and (b) with CAD. (a, b) R1 = reader 1, R2 = reader 2, R3 = reader 3, R4 = reader 4, R5 = reader 5, R6 = reader 6. All six observers achieved an increase in the area under the ROC curve, Az, with CAD.

View larger version (28K): [in a new window]   Figure 8. Average ROC curve obtained from the average a and b parameters of the six individual ROC curves for each of the four reading conditions. An improved ROC curve was achieved with CAD in both the single-view and two-view reading experiments.

View larger version (31K): [in a new window]   Figure 9. PPV as a function of the false-negative fraction derived from the ROC curves for the six observers (Fig 7). The PPV was predicted for a population of masses in which biopsy was likely to be performed under current clinical criteria and by assuming the prevalence of malignant masses to be 25%. R1 = reader 1, R2 = reader 2, R3 = reader 3, R4 = reader 4, R5 = reader 5, R6 = reader 6.

View larger version (17K): [in a new window]   Figure 10a. Histograms illustrate the confidence ratings of reader 5 obtained by reading 76 two-view mammograms (a) without CAD and (b) with CAD. The specificity of reader 5 at 100% sensitivity would increase from 5% (two of 37 masses) without CAD to 68% (25 of 37 masses) with CAD if an appropriate decision threshold were chosen.

View larger version (17K): [in a new window]   Figure 10b. Histograms illustrate the confidence ratings of reader 5 obtained by reading 76 two-view mammograms (a) without CAD and (b) with CAD. The specificity of reader 5 at 100% sensitivity would increase from 5% (two of 37 masses) without CAD to 68% (25 of 37 masses) with CAD if an appropriate decision threshold were chosen.