Vision-based Self-supervised Representation: from Foundational Innovations to Clinical Diagnostics

Abstract

Supervised learning in computer vision has achieved remarkable success but remains constrained by its dependence on extensive labeled datasets, which introduce annotation costs, scalability limitations, label noise, and dataset bias. Self-supervised learning has recently emerged as a promising alternative, enabling systems to learn meaningful representations from unlabeled data. This thesis presents novel contributions to self-supervised visual representation learning and its applications in medical imaging, addressing fundamental challenges in representation learning while demonstrating practical solutions for healthcare technology. The work is organized into two primary parts. The first ..