Abstract

Breast cancer is the most commonly diagnosed malignancy among women worldwide and remains a leading cause of cancer-related mortality. Early and accurate diagnosis is essential for improving survival while minimizing unnecessary interventions. Conventional imaging modalities including mammography, ultrasound (US), and magnetic resonance imaging, provide critical structural information but often suffer from limited specificity, particularly in dense breast tissue, resulting in high false-positive rates and avoidable biopsies. Diffuse optical tomography (DOT) is a functional imaging modality that noninvasively measures tissue hemoglobin concentration and oxygenation using near-infrared light, offering complementary physiological insight into tumor angiogenesis and metabolism. Despite its promise, clinical translation of DOT has been hindered by the ill-posed nature of the inverse problem, susceptibility to measurement artifacts, limited spatial resolution, and reliance on time-consuming, expert-driven data processing. This dissertation presents a comprehensive framework to advance ultrasound-guided diffuse optical tomography (US-DOT) toward fully automated, robust, and clinically practical breast cancer diagnosis through the integration of deep learning and multimodal data fusion. First, an automated US-guided DOT processing and diagnostic pipeline is developed that incorporates motion detection, reference selection, mismatch identification, outlier rejection, and a two-stage classification strategy. By minimizing user interaction and automating key preprocessing steps, this pipeline reduces total processing time from tens of minutes to near real time while preserving diagnostic accuracy comparable to expert manual workflows. Importantly, the two-stage design enables rapid identification of clearly benign lesions, reducing the need for full DOT reconstruction and improving clinical efficiency. Second, a physics-aware deep learning framework, termed APU-Net, is introduced to enhance DOT image quality. APU-Net integrates attention mechanisms with a U-Net architecture and incorporates physical constraints derived from DOT forward modeling. Trained using a combination of simulated and phantom data, the model effectively suppresses reconstruction artifacts, improves depth fidelity, and enhances lesion contrast in clinical DOT images, addressing long-standing limitations of conventional reconstruction methods. Third, a convolutional–transformer–based multimodal fusion model is proposed to jointly analyze high-resolution US images and lower-resolution DOT hemoglobin maps for lesion classification. By learning complementary structural and functional representations across spatial scales, this fusion strategy improves diagnostic performance relative to single-modality approaches and demonstrates potential to reduce unnecessary biopsies when compared with conventional imaging assessments. Collectively, this work establishes an end-to-end, deep learning–enabled US-guided DOT framework that enhances robustness, efficiency, and diagnostic utility. The results support the clinical feasibility of automated functional–structural imaging as an adjunct to standard breast imaging, with implications for improving diagnostic specificity, reducing patient burden, and enabling more personalized breast cancer care.

Committee Chair

Quing Zhu

Committee Members

Abhinav K. Jha; Adam Q. Bauer; Christine M. O’Brien; Hong Chen

Degree

Doctor of Philosophy (PhD)

Author's Department

Biomedical Engineering

Author's School

McKelvey School of Engineering

Document Type

Dissertation

Date of Award

4-29-2026

Language

English (en)

Download

Available for download on Tuesday, June 15, 2027

Share

COinS