January 27, 2020 –
Title: Deep Learning Systems for Automated Lesion Detection, Segmentation, and Classification in Mammography
Day/Time: Monday, Jan 27, 2020, 9:30 am
Location: MCHU 302
Breast cancer is the second leading cause of cancer deaths among women in the USA. Mammography is the preferred screening tool for breast cancer and accounts for the greatest contribution to the early detection of breast cancer. The detection of breast masses in mammogram (MG) images using deep learning (DL) systems is a challenging task due to the varying sizes, shapes, and textures of masses.
In this thesis, we propose a novel DL network called residual attention UNet (RAU-Net), the network pays attention to small lesions, and shows superior performance compared to the other state-of-the-arts DL models in detecting and segmenting masses, especially for heterogeneously dense and dense MG images. The proposed RAU-Net model achieves a mean dice coefficient index of 0.98 and mean intersection over union of 0.94. We propose a DL residual network for classification of MG images into benign and malignant that achieved accuracy of 0.95, and AUC of 0.98.
We also propose a one-shot multi-input Siamese network that learns features from previous and current year MG images of the same patient to give a better assessment for current year MG images. The detection of mass tumors in dense tissues and, more generally, in dense breasts is often considered more challenging due to the similar visual aspects of normal and abnormal dense tissues.
In this thesis, we present a training algorithm that we used to train various kinds of U-Net networks such as RCNN-UNet, AU-Net, RAU-Net, and UNet++ to generate density attention masks that automatically pays attention and gives more weight to tumors in dense regions of MG images. To train and test our models, we collected and pre-processed MG images that come with different resolutions from public repositories and MG images from UCONN health center. In conclusion, we proposed DL systems for lesion detection, segmentation, and classification in mammography that can aid radiologists and serve as a second eye for them.