Improved diagnosis of breast microcalcification clusters

点击翻译按钮获取中文摘要

The purpose of this study is to develop computer models to increase the specificity of breast biopsy by improving the diagnosis of microcalcifications in mammograms. Probably benign cases may undergo short-term follow-up in lieu of biopsy. To accomplish these goals, computer models will merge together three complementary sources of diagnostic information: radiologist-extracted mammographic findings, patient history data, and computer-extracted I features from local histogram thresholding applied to digitized mammograms. The specific aims of the proposed study are to: (1) Optimize predictive models to identify probably benign microcalcification clusters based upon radiologist-extracted findings and history data. (2) Design new predictive models based upon computer-extracted features from local histogram analysis. (3) Construct and evaluate a unified malignancy predictor which combines the different information provided by the separate models. In preliminary studies, an ANN identified probably benign cases of clustered microcalcifications using BI-RADS findings and history data as inputs to the model. In addition, a local histogram thresholding technique was used to segment microcalcification clusters, and a rule- based system eliminated typically benign clusters. The immediate benefit of this proposal is a computer-based decision aid for the diagnosis of mammographically suspicious lesions with microcalcification clusters. These cases account for 40% of breast biopsies and are arguably the most difficult category to characterize for radiologists and computer models alike. Improvements in diagnostic accuracy for these cases of microcalcification clusters will thus have an important and immediate impact.