Quantitative assessment of breast dense tissue on mammograms

This paper presents a method to automatically segment dense tissue from mammography images. The method uses unsupervised learning and multiple levels of detail (LoD) to adapt itself to various image characteristics and generate robust segmentation result, which can potentially be of great help in surveillance breast health and detecting breast cancer at early stage. Multiple LoD used in the method include 1. Initial entropy maximum based thresholding (low LoD); 2. FCM based ¿soft-threshold¿ estimation (mid LoD); 3 Pixel-wise dense tissue feature evaluation (high LodD). The performance validation, based on 220 cases from 6 different mammographic data sets, shows a strong correlation between computer and clinical BI-RADS ratings.