Automated classification of breast lesions on digital mammograms

While the sensitivity of mammography is approximately 85%, the positive predictive value (PPV) is only 10%-40% [1, 2]. Thus, the majority of breast biopsies are performed on women who do not have breast cancer. Biopsies are expensive, invasive, and traumatic to the patient. The relatively high level of benign biopsies is often quoted as one of the "risks" associated with screening mammography. The low PPV of mammography is due in part to the difficulty in distinguishing benign from malignant lesions on the basis of mammograms [1, 3], and to the desire of the radiologist that a cancer not be missed [2]. We believe, by providing the radiologist with the results of quantitative analyses of breast lesions, that the PPV of mammography can be increased without a loss in sensitivity. In this paper, we describe two automated techniques for classifying breast lesions as benign or malignant. One technique is for masses [4-6] and the other is for clustered microcalcifications [7-9]. Both schemes operate in a similar manner: the computer automatically extracts features of the lesion, and then these features are merged using either an artificial neural network (ANN) or a combination of rule-based and ANN approach. The output of the scheme is the lesion's likelihood of malignancy. Both techniques use screen-film mammograms digitized to 100-micron pixel size and 10-bit grey scale resolution.