Deep learning for computer aided diagnosis in mammography

Abstract

Breast cancer is the most frequently diagnosed and the leading cause of the cancer death for women worldwide. Mammography is an intensively used breast screening modality for early diagnosis. The manual inspection by radiologists is tedious, subjective, and prone to errors. A computer-aided system is thus desired as an alternative to double-reading. Recent advances in deep learning have spiked tremendous interests in automatic clinician diagnosis based on medical images. Early detection has proven to be critical to give patients the best chance of recovery and survival. Advanced computer-aided diagnosis systems are expected to have high sensitivities and low positive rates and how best to provide accurate diagnosis results is an important topic in the current computer-aided diagnosis research. In this thesis, a series of mammography image processing methods are proposed, including automatic abnormality contouring, and cancer diagnosis.

First, we will show how to employ the probabilistic graph and deep learning to improve mammogram mass segmentation. Dice similarity coefficients (DSC) were used to evaluate the contours. The results show a good agreement between clinical contours and the proposed segmentation algorithm CRU-Net, the DSCs are found to be state-of-art on two public mammography datasets, as 92.17\% for DDSM and 93.66 \% for INbreast.

Then a novel algorithm was developed for weakly labelled mammogram data for breast cancer diagnosis. By efficiently integrating automatic segmentation results in deep learning classification model, an improved diagnosis performance was noticed. Area under the receiver operating characteristic curve (AUC) and accuracy are utilized for assessment, in 90\% cases were found to between 85\% and 92\% AUC.

Lastly, an innovative method, in which the graph embedding was incorporated into deep learning in medical image analysis. Furthermore, an adversarial augmentation method was proposed to generate both on- and off-manifold neighboring instances. The effectiveness of graph embedding and adversarial augmentation for mammogram cancer classification were evaluated. Experiment results have shown that the proposed algorithm \textsc{DiagNet} achieves the state-of-art results in the INbreast dataset.