Automatic Breast Mass Detection and Classification for Mammograms with Deep Learning

Breast cancer has become one of the most concerning cancers that are well known for its high incidence rate and mortality. Integration of novel deep learning techniques in computer aided systems for quick diagnosis of breast cancer caused by breast mass is of great importance. In thesis, I aimed at developing a breast mass detection and classification system for mammography images based on deep learning techniques.The developed system is comprised of three modules including pre-processing, breast mass detection, and breast mass classification. The pre-processing module mainly focuses on standardizing the input images for following computation. The main contribution within pre-processing module is that I developed a novel attention mechanism called global channel attention module, which helps the deep learning network to learn inter-channel attention and thus improves the segmentation performance of deep learning networks. To develop a breast mass detection module with higher detection performance but with lower training cost, I then developed a new patch-based breast mass detection system, within which I improved the patch extraction algorithm and developed a new false positive suppression algorithm. Thanks to proposed algorithms, the developed breast mass detection system can achieve the state-of-the-art performance. In the newly developed breast mass classification module, I proposed to feed classifiers with fused deep learning features. By doing so, the overall classification performance can be improved thanks to better usage of information. Moreover, the training efficiency of the novel classifier is much higher than the back propagation based learning algorithms thanks to the novel architectures introduced. The experiments on public datasets showed high performance of the three developed modules against the state-of-the-art methods while some of them showed even higher performance than the state-of-the-art methods.