Classification of mammographic tissue using shape and texture features

Abstract

The authors have performed a pilot study on the classification of regions of interest (ROI) containing normal tissue, biopsy-proven malignant masses, and biopsy-proven microcalcification (MCC) clusters using a mix of shape and texture features. Shape features included size, translation, and rotation invariant moments. Texture features included fractal-based features and spatial gray level dependence (SGLD) matrix features. The entropy was also computed for each ROI. The type of classifier used was a neural network based on the ALOPEX training algorithm. Best results using a database of 40 normal, 32 mass, and 20 MCC ROIs for training and 5 normal, 5 mass, and 4 MCC ROIs for testing were obtained using texture features and a binary tree neural network which splits the classification of the three types of tissue into two steps. Classification between normal ROIs and abnormal (mass+MCC) ROIs reached 95% training and 100% testing and classification of mass and MCC ROIs reached 98% training and 100% testing.