Hui Li PhD, Presenter: Nothing to Disclose

Maryellen L. Giger PhD, Abstract Co-Author: Stockholder, Hologic, Inc Shareholder, Quantitative Insights, Inc Royalties, Hologic, Inc Royalties, General Electric Company Royalties, MEDIAN Technologies Royalties, Riverain Medical Royalties, Mitsubishi Corporation Royalties, Toshiba Corporation Researcher, Koninklijke Philips Electronics NV Researcher, U-Systems, Inc

Li Lan, Abstract Co-Author: Nothing to Disclose

Yading Yuan, Abstract Co-Author: Nothing to Disclose

Charlene A. Sennett MD, Abstract Co-Author: Nothing to Disclose

To investigate how different computerized tumor segmentation methods affect classification performance of computer-aided diagnosis on full-field digital mammograms. We performed two different lesion segmentation methods on a large clinical FFDM dataset, and evaluated the corresponding CADx performance in terms of distinguishing malignant and benign lesions.

Evaluation was performed on 287 biopsy-proven lesions, including 148 malignant and 139 benign lesions. Lesion margins were delineated by an expert breast radiologist and were used as the truth in the lesion-segmentation evaluation. Our CADx method includes: 1) automatic lesion segmentation from parenchymal background; 2) automatic feature extraction; 3) automatic feature selection; and 4) automatic probability of malignancy estimation using a Bayesian neural network (BANN). Two different computerized lesion segmentation methods, region growing and active contour methods, were performed on each lesion and evaluated using ROC analysis with area under the ROC curve as the performance index.

At an overlap threshold of 0.4, 86% and 69% of images were correctly segmented with active contour and region growing methods, respectively. The spiculation and radial gradient index features extracted from the active contour segmented tumors performed significantly better than those extracted from the region growing segmented tumors in the classification task (p-value=0.0023). The BANN classifier, in a round-robin evaluation, yielded AUC values of 0.85 (SE=0.02) and 0.81 (SE=0.02) in the task of distinguishing malignant from benign lesions segmented by active contour and region growing methods, respectively (p-value=0.05).

The initial segmentation of breast tumors is an essential component within CADx systems and should include evaluations in terms of both overlap and classification performance. Identification of the more affected tumor features, i.e., spiculation, can direct the next stage of segmentation development.

Ultimate clinical usage of CADx will depend on system performance, which is dependent on initial tumor segmentation. Careful evaluation of the segmentation is thus critical to overall performance.

Li, H, Giger, M, Lan, L, Yuan, Y, Sennett, C, Effect of Breast Tumor Segmentation on CADx Performance Evaluated on a Large Clinical FFDM Dataset.  Radiological Society of North America 2012 Scientific Assembly and Annual Meeting, November 25 - November 30, 2012 ,Chicago IL. http://archive.rsna.org/2012/12043886.html