Towards High-Fidelity Prostate Tissue Characterization and Cancer Detection with Micro-Ultrasound and Deep Learning

Abstract

Machine learning has revolutionized medical imaging by transferring and distilling vast amounts of expert human knowledge into automated systems capable of detecting complex patterns in medical images. This thesis focuses on advancing machine learning techniques for one of the most critical applications in medical imaging: detecting prostate cancer from ultrasound images and enabling targeted biopsy. Ultrasound-guided systematic biopsy, followed by pathological grading, is the current standard-of-care for prostate cancer diagnosis. However, conventional transrectal ultrasound (TRUS) lacks the ability to differentiate between malignant and benign tissues, leading to suboptimal biopsy targeting and high false-negative rates. High-resolution micro-ultrasound (micro-US) has emerged as a promising imaging modality that provides three times higher resolution than conventional ultrasound, offering improved visualization of prostate tissue microstructures. Despite its potential, micro-US remains underutilized in automated machine-enhanced cancer detection due to the challenges of limited labeled data, weak pathology annotations, and distribution shifts across clinical centers. To address these challenges, this thesis proposes a series of deep learning-based approaches aimed at improving the reliability and robustness of prostate cancer detection from micro-US. First, an uncertainty-aware deep learning model is introduced to provide confidence estimates for predictions, enabling improved biopsy targeting by reducing unreliable classifications. Second, a self-supervised learning framework is developed to mitigate the scarcity of labeled data by leveraging large amounts of unlabeled micro-US images. Third, a multiple instance learning-based approach is introduced to overcome the issue of weak pathology annotations by learning from biopsy core-level labels while preserving fine-scale cancer localization. Finally, a test-time adaptation technique is proposed to enhance model generalizability across clinical centers with distribution shifts. Through extensive experiments on multi-center datasets, the proposed methods demonstrate significant improvements in prostate cancer detection accuracy, model reliability, and generalization. The findings of this thesis contribute to the advancement of AI-driven prostate cancer diagnosis, bringing micro-US one step closer to real-time, targeted biopsy guidance in clinical practice.