A computer‐aided diagnostic scheme for classification of clustered microcalcifications in mammograms

Abstract

The hypothesis of this dissertation is that computer‐aided diagnosis (CAD) can improve radiologists' diagnostic performance in breast cancer diagnosis. An automated computerized classification scheme was developed to differentiate malignant from benign clustered microcalcifications. This computer scheme used an artificial neural network (ANN) to analyze eight computer‐extracted image features based on number, size, uniformity, and shape of individual microcalcifications, and on size and shape of a cluster. The features correlated qualitatively with radiologists' perceptual experience. The performance of this computer scheme was evaluated on two independent databases of standard‐view mammograms digitized at 0.1‐mm pixel size: database A contained mammograms from 53 patients with biopsies and database B contained mammograms from 104 patients in a consecutive biopsy series. The computer scheme performed better than radiologists on both databases (p=0.03 and p<0.0001, respectively). An observer performance study was conducted using database B. Ten radiologists read the original standard‐ and magnification‐view mammograms with and without the aid of our computer scheme. The performance of the radiologists improved significantly (p<0.0001) when the computer aid was used: on average, each radiologist recommended 6.4 more biopsies for malignant tumors (p=0.0006) and 6.0 fewer biopsies for benign lesions (p=0.003). We concluded that our computer scheme classifies malignant and benign clustered microcalcifications more accurately than radiologists, and that it can be used to reduce the number of biopsies for benign lesions while maintaining or improving sensitivity for breast cancer diagnosis.