Date of Award:


Document Type:


Degree Name:

Doctor of Philosophy (PhD)


Computer Science

Committee Chair(s)

Xiaojun Qi


Xiaojun Qi


Vicki Allan


David Brown


John Edwards


Haitao Wang


Breast cancer is a great threat to women’s health. Breast ultrasound (BUS) imaging is commonly used in the early detection of breast cancer as a portable, valuable, and widely available diagnosis tool. Automated BUS image analysis can assist radiologists in making accurate and fast decisions. Generally, automated BUS image analysis includes BUS image segmentation and classification. BUS image segmentation automatically extracts tumor regions from a BUS image. BUS image classification automatically classifies breast tumors into benign or malignant categories. Multi-task learning accomplishes segmentation and classification simultaneously, which makes it more appealing and practical than an either individual task. Deep neural networks have recently been employed to achieve better image segmentation and classification results than conventional approaches. In addition, attention mechanisms are applied to deep neural networks to make them focus on the important parts of the input to improve the segmentation and classification performance. However, BUS image segmentation and classification are still challenging due to the lack of public training data and the high variability of tumors in shape, size, and location.

In this dissertation, we introduce three different deep learning architectures with attention mechanisms, each of which aims to address the drawbacks of their peers and evaluate their performance in terms of segmentation and classification accuracy on two public BUS datasets. First, we propose a Multi-Scale Self-Attention Network (MSSA-Net) for BUS image segmentation that can be trained on small BUS image datasets. We design a multi-scale attention mechanism to explore relationships between pixels to improve the feature representation and achieve better segmentation accuracy. Second, we propose a Multi-Task Learning Network with Context-Oriented Self-Attention (MTL-COSA) to segment tumors and classify them as benign or malignant automatically and simultaneously. We design a COSA attention mechanism that utilizes segmentation outputs to estimate the tumor boundary, which is treated as prior medical knowledge, to guide the network to learn contextual relationships for better feature representations to improve both segmentation and classification accuracy. Third, we propose a Regional-Attentive Multi-Task Learning framework (RMTL-Net) for simultaneous BUS image segmentation and classification. We design a regional attention mechanism that employs the segmentation output to guide the classifier to learn important category-sensitive information in three regions of BUS images and fuse them to achieve better classification accuracy. We conduct experiments on two public BUS image datasets to show the superiority of the proposed three methods to several state-of-the-art methods for BUS image segmentation, classification, and Multi-task learning.