Computer-aided diagnosis has great potential to improve performance of the detection and classification of abnormalities on breast ultrasound. Our goal is to develop a computerized tool that detects suspicious areas and distinguishes among false-positive
detections, benign lesions, and malignancies. One-step classification into 3 categories is a largely unexplored territory with many possibilities.
The computerized scheme first identifies potential lesions based on expected lesion shape and margin characteristics. Our main focus here, however, is the subsequent classification of the potential lesions into 3 categories. For this purpose, we use a 3-way Bayesian neural net (BNN) based on extracted image features of the lesion candidates.
The method was tested on a database of 858 cases (1832 images) consisting of complex and simple cysts, benign solid lesions, and malignant lesions. In order to verify whether performance conforms to expectations, the output of a 3-way classifier ("A" vs. "B" vs. "C") can be projected onto that of two 2-way classifiers ("A" vs. "B or C", and "A or B" vs. "C"). We compared the projected performance of the 3-way classifier to two specifically trained 2-way classifiers. The first task was to distinguish cancer from all other lesion candidates, and the second was to distinguish actual lesions from false-positive detections. For these tasks, the performance of the 2-way classifiers and the projected performance of the 3-way classifier were indistinguishable based on calculated means and standard deviations of ROC area (Az). For example, in round robin analysis the average Az values obtained with both approaches were 0.92 and 0.83, for the two tasks, with standard deviations of 0.006 and 0.010, respectively. The potential of 3-way classification is illustrated graphically through the estimated probability density functions of the three truth categories. We have implemented a promising computerized scheme for detection and subsequent one-step 3-way classification of breast lesions on ultrasound images. The method was tested on an extensive database. The main challenge is the development of an objective evaluation method of performance such as the equivalent of ROC analysis for the 2-class problem.