Efficient and Robust Self-Supervised Learning for Deep Learning-Based Healthcare Applications

Abstract

As healthcare increasingly relies on deep learning for medical imaging, a critical challenge arises: the scarcity of labeled data due to expensive and time-consuming manual clinical annotation. This thesis addresses the mismatch between deep learning's heavy data demands and clinical data scarcity by exploring Self-Supervised Learning (SSL). SSL learns meaningful representations from unlabeled data, significantly reducing dependency on extensive annotations by leveraging inherent data structures and relationships.

The primary objective of this research is to develop novel SSL methodologies tailored to distinct healthcare analysis tasks, maximizing the efficient use of limited data across multiple scales and modalities. This is demonstrated through three domain-specific innovations:

1. Histopathology: A novel SSL framework leverages the multi-resolution nature of whole-slide images to enable hierarchical representation learning. This effectively captures both global tissue organization and fine-grained cellular details.

2. Dermatology: To mirror clinical workflows, SSL is customized with pretext tasks that align multi-modal representations (clinical and dermoscopic images) and encode inter-label dependencies for complex diagnostic predictions.

3. Remote Physiological Measurement: To extract subtle spatiotemporal signals from facial videos, SSL is extended with physiology-aware temporal and spatial augmentations. This preserves periodic signal integrity while efficiently suppressing noise.

Through these investigations, this thesis demonstrates that SSL can be successfully adapted to exploit domain-specific data characteristics—such as multi-resolution hierarchies, multi-modal complementarity, and spatiotemporal dynamics. Ultimately, this research introduces a robust, general SSL framework that significantly reduces annotation requirements while consistently achieving state-of-the-art predictive performance across diverse healthcare applications.