Date of Award:


Document Type:


Degree Name:

Doctor of Philosophy (PhD)


Computer Science

Committee Chair(s)

Heng-Da Cheng


Heng-Da Cheng


Lie Zhu


Haitao Wang


Curtis Dyreson


Vicki Allan


Breast cancer frequently occurs in women over the world. It was one of the most serious diseases and the second common cancer among women in 2019. The survival rate of stages 0 and 1 of breast cancer is closed to 100%. It is urgent to develop an approach that can detect breast cancer in the early stages. Breast ultrasound (BUS) imaging is low-cost, portable, and effective; therefore, it becomes the most crucial approach for breast cancer diagnosis. However, BUS images are of poor quality, low contrast, and uncertain. The computer-aided diagnosis (CAD) system is developed for breast cancer to prevent misdiagnosis.

There have been many types of research for BUS image segmentation based on classic machine learning and computer vision methods, e.g., clustering methods, thresholding methods, level set, active contour, and graph cut. Since deep neural networks have been widely utilized in nature image semantic segmentation and achieved good results, deep learning approaches are also applied to BUS image segmentation. However, the previous methods still suffer some shortcomings. Firstly, the previous non-deep learning approaches highly depend on the manually selected features, such as texture, frequency, and intensity. Secondly, the previous deep learning approaches do not solve the uncertainty and noise in BUS images and deep learning architectures. Meanwhile, the previous methods also do not involve context information such as medical knowledge about breast cancer. In this work, three approaches are proposed to measure and reduce uncertainty and noise in deep neural networks. Also, three approaches are designed to involve medical knowledge and long-range distance context information in machine learning algorithms. The proposed methods are applied to breast ultrasound image segmentation.

In the first part, three fuzzy uncertainty reduction architectures are designed to measure the uncertainty degree for pixels and channels in the convolutional feature maps. Then, medical knowledge constrained conditional random fields are proposed to reflect the breast layer structure and refine the segmentation results. A novel shape-adaptive convolutional operator is proposed to provide long-distance context information in the convolutional layer. Finally, a fuzzy generative adversarial network is proposed to reduce uncertainty. The new approaches are applied to 4 breast ultrasound image datasets: one multi-category dataset and three public datasets with pixel-wise ground truths for tumor and background. The proposed methods achieve the best performance among 15 BUS image segmentation methods on the four datasets.