Open Collections

Abstract

Prostate cancer (PCa) is a highly prevalent disease with a heterogeneous clinical course, where accurate detection of clinically significant cancer is critical for guiding treatment decisions and improving patient outcomes. Transrectal ultrasound (TRUS) remains the most widely used imaging modality for prostate biopsy guidance due to its accessibility and real-time capabilities, yet its limited tissue contrast and reliance on systematic sampling result in suboptimal sensitivity and specificity. This thesis investigates advanced learning-based approaches to enhance ultrasound-only prostate cancer detection by leveraging temporal information, robust representation learning, and cross-modal knowledge transfer under weak supervision. First, this work explores temporal enhanced ultrasound (TeUS), which analyzes time-series radio-frequency signals to capture subtle tissue dynamics beyond static imaging. A self-supervised prototype learning framework is introduced to improve robustness to noisy, biopsy-level labels by aligning spatial and temporal representations prior to supervised training. This approach mitigates label uncertainty and improves downstream detection performance. Building on this foundation, a spatio-temporal learning framework is then proposed that integrates global spatial features, fine-grained temporal signatures, and clinical metadata within a unified architecture. A hybrid loss function and progressive training strategy are employed to further enhance resilience to weak and noisy supervision. Finally, this thesis introduces a novel histopathology-guided learning framework for micro-ultrasound (micro-US) prostate cancer assessment. By distilling knowledge from a pretrained whole-slide histopathology teacher model into a micro-US student encoder using weakly paired data, the proposed method enables the imaging model to learn pathology-informed representations without requiring spatial registration. An attention-based multiple instance learning mechanism addresses disparities in scale and anatomical coverage between modalities. Together, the methods presented in this thesis demonstrate consistent improvements in prostate cancer detection across multiple ultrasound settings and establish scalable, ultrasound-only solutions that reduce reliance on costly imaging modalities. This work advances the clinical viability of ultrasound-based prostate cancer detection and provides a foundation for more precise, robust, and accessible biopsy guidance.