PhD Oral Exam - Bahareh Behboodi, Electrical and Computer Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Ultrasound (US) is one of the most widely used imaging modalities in both diagnostic and surgical settings due to its affordability, safety, and non-invasive nature. Research in US imaging can be broadly divided into two main categories: pre-image formation and post-image formation. Pre-image formation research focuses on advancements in hardware, beamforming techniques, and image reconstruction methodologies. These studies aim to improve the initial creation of ultrasound images. Post-image formation research, which is the primary focus of this Ph.D. thesis, involves developing image processing algorithms to enhance visualization and interpretation. This includes techniques such as image enhancement, segmentation, classification, and more. Despite the expanding applications of US imaging, interpreting US images remains challenging due to their inherently low quality. Even experienced physicians find manual analysis time-consuming and complex. To address these challenges, researchers have increasingly turned to automatic image processing methods, particularly deep learning (DL) algorithms, to support radiologists and accelerate the diagnostic and surgical process. Although DL models have shown strong performance, they require customization for specific applications and often demand large training datasets. Unfortunately, acquiring large, labeled datasets in medical imaging is particularly difficult due to privacy concerns, posing a significant barrier to their widespread implementation.

To overcome these obstacles, this thesis proposes several DL-based methods designed to address these challenges. Specifically, in Chapters 2, 3, and 4, we introduce novel segmentation techniques that are effective even with limited data. After the Introduction Chapter, we develop an innovative classification method for breast lesion classification using a small dataset in Chapter 5. In Chapter 6, we integrate our expertise in segmentation, classification, and regression to analyze real clinical datasets with limited patient US images. In Chapter 7, we present a publicly available brain US dataset with tumor annotations and present concluding remarks in Chapter 8. These contributions aim to advance the field of US image processing, bringing it closer to practical clinical application.