Texture-Based Breast Cancer Prediction in Full-Field Digital Mammograms Using the Dual-Tree Complex Wavelet Transform and Random Forest Classification

In this paper we describe a novel methodology for texture-based breast cancer prediction in full-field digital mammograms. Our method employs the Dual-Tree Complex Wavelet Transform for texture-based image analysis and representation, and Random Forest classification for discriminative learning and breast cancer prediction. We assess the ability of our method to identify women with breast cancer using raw images, processed images and VolparaTM density maps of two case-control datasets. We also investigate whether different regions of the breast exhibit different predictive power with respect to breast cancer. The best results are obtained using the processed images of a case-control dataset consisting of 100 cancers and 300 controls, where we achieve an area under the ROC curve of 0.74 for a texture model based on the whole breast and an equal area under the ROC curve when the most predictive regional model is used.