Meta-Learning for Rapid Adaptation and Generalization in Medical Image Classification

Abstract

This thesis investigates the application of meta-learning approaches for rapid adaptation in medical image classification, with a particular focus on prostate cancer and colorectal cancer detection. Classical deep learning models require large annotated datasets and extensive retraining when applied to new tasks, which is a critical limitation in medical imaging where labeled data is scarce. Meta-learning, or “learning to learn”, enables models to acquire transferable knowledge across tasks and adapt efficiently to new classification problems with minimal labeled examples. In parallel, self-supervised learning (SSL) is explored as an independent pretraining strategy to extract rich and generalizable visual representations from unlabeled medical images, without relying on manual annotations. We evaluate state-of-the-art meta-learning algorithms including Model-Agnostic Meta-Learning (MAML) and Prototypical Networks as well as SSL-based approaches, on histopathological image datasets for prostate and colorectal cancer classification. Our proposed approaches demonstrate significant improvements in few-shot adaptation performance compared to conventional transfer learning baselines.