Breast tumor detection in double views mammography based on extreme learning machine

Abstract

Mammography is one of the most important methods for breast tumor detection, while existing computer-aided diagnosis (CAD) technology based on single-view mammograms ignores the contrastive feature between medio-lateral oblique (MLO) and cranio-caudal (CC) views, and CAD technology based on double-view overlooks features of single views. But in clinical environment, radiologists not only read both CC view images and MLO view images individually, but also contrast these two types of views to diagnose each case. Therefore, to simulate diagnosis process of radiologists, in this paper, a fused feature model which blends features of single views with contrastive features of double views is proposed. The fused feature model is optimized by means of feature selection methods. Then, a CAD detection method based on extreme learning machine, a classifier with wonderful universal approximation capability, is proposed to improve the effectiveness of breast tumor detection by applying the optimum fused feature. The effectiveness of proposed method is verified by 222 pairs of mammograms from 222 women in Northeast China through the complete experiment.