Computer-aided detection of microcalcifications in mammography using deep neural networks

Breast cancer accounts for 22.9% of diagnosed cancers and it is the third highest cause of cancer-related mortality worldwide. According to the National Cancer Institute, early diagnosis of breast cancer is of utmost importance in effective treatment for increasing the survival rate for this disease. The diagnosis of breast cancer is mainly performed by periodically screening the breast region by human experts to be able to detect a potential threat at an early stage. However, the screening outcomes are highly detailed images that might be confusing in some cases considering the capabilities of the human visual system. For this reason, advanced computer-based systems are developed in order to support the radiologists in recent years. This collaboration is found particularly useful to assist radiologists in diagnosing subtle abnormalities appear in the screening that might not be obvious to the eye otherwise. In this thesis, we investigated the power of deep learning algorithms in detecting microcalcification lesions which are one of the most common signs of breast cancer. We propose a performance comparison between different kinds of neural network architectures (SqueezeNet and Inception ResNet) together with a powerful data augmentation operation that allows us to train these networks with publicly available datasets. After analyzing the performance of these neural networks, we further inspected the transfer learning capabilities of these solutions to be applied on full-field digital mammography. The novelty of our approach relies on the efficiency of using the publicly available datasets and the power of feature extraction capabilities of both relatively smaller and very deep neural network architectures in microcalcification detection task. The proposed pipeline is capable of detecting microcalcifications both in screen film and full-field digital mammograms with AUC values of 0.99 and 0.98 respectively. To our knowledge, this performance represents the state of the art in microcalcification detection in the literature which consistently beats human performance at the same time.