Method, system and computer program product for breast density classification using parts-based local features

Abstract

An automated content-based image retrieval method, a system and a computer program product for the classification of breast density from mammographic imagery. Raw digital mammogram images taken of patients are initially pre-processed to remove noise and enhance contrast, then subjected to pectoral muscle segmentation to produce region of interest (ROI) images. The ROI images are then decomposed using non-negative matrix factorization (NMF), where a non-negative sparsity constraint and reconstruction quality measures are imposed on the extracted and retained first few NMF factors. Based on the retained NMF factors, kernel matrix-based support vector machines classify the mammogram images binomially or multinomially to breast density categories. Methods of assessing and comparing the NMF-based breast classification method to principal component analysis or PCA-based methods are also described, and the NMF-based method is found to achieve higher classification accuracy and better handling of invariance in the digital mammogram images because of its parts-based factorization.