Computer-aided classification of breast masses in ultrasonic B-scans using a multiparameter approach

Classification of breast masses in ultrasonic B-scan images is undertaken using a multiparameter approach. The parameters are generated on the basis of a non-Rayleigh statistic model of the backscattered envelope from the breast tissue. They can be computed automatically with minimal clinical intervention once the location of the mass is known. A new discriminant is developed that combines these parameters linearly. It is seen that this new discriminant performs classification of masses into benign or malignant better than the classification by any one of the individual parameters. The data set studied consisted of 99 cases (70 patients with benign masses and 29 patients with malignant masses). The areas under the receiver operating characteristic (ROC) curves (Az) and statistical attributes of the areas were studied to establish the enhancement in performance. The Az value after combining all the parameters was found to be 0.8701. Upon combining this parameter with the level of suspicion (LOS) scores of a radiologist, the performance is further enhanced with an area under the (empirical) ROC of 0.94 having an operating point at a sensitivity of 0.965 and specificity of 0.87. It is suggested that this automated approach may hold promise as a means of classifying breast masses.