Breast mass classification on sonographic images on the basis of shape analysis

To evaluate the efficiency of novel shape features for classification of benign and malignant sonographic breast masses.

Mass regions were extracted from the region of interest (ROI) sub-image by applying a segmentation algorithm based on the level set method. Six features (difference area with five features of mass pixel number viewed at different angles) were then extracted for further classification. A multilayered perceptron neural network (MLP) classifier was used to classify breast mass. The leave-one-case-out procedure was used on a database of 81 pathologically proved breast sonographic images of patients (47 benign cases and 34 malignant cases) to evaluate our method.

The classification results showed overall accuracy was 93.83%, sensitivity 91.18%, specificity 95.74%, positive predictive value 93.94%, and negative predictive value 93.75%.

The experimental results showed that this diagnostic system with the features proposed can improve the positive rate of biopsies, provide a second opinion for physicians, and be used as a useful tool for mass classification.