A review of the evolution and challenges of few-shot medical image semantic segmentation

Image segmentation, as a fundamental task in computer vision, aims to achieve pixel-level delineation between target regions and background. In the field of medical image analysis, traditional segmentation methods heavily rely on large-scale, high-quality annotated datasets. However, expert-level annotations for medical images require substantial professional human resources ( e.g ., radiologists performing slice-by-slice delineation of lesions) and face challenges such as cross-institutional data heterogeneity. To address this data bottleneck, few-shot medical image semantic segmentation has emerged as a promising paradigm. By leveraging strategies such as meta-learning, metric learning, and other related approaches, these approaches enable models to accurately segment the target region and achieve effective generalization to novel categories using minimal annotated samples (typically ≤15 instances per class), thereby significantly reducing dependence on annotated data. This article systematically explores foundational concepts such as medical image segmentation methods, benchmark datasets, performance evaluation metrics, and main approaches for few-shot medical image segmentation. Distinct from existing review literature, our work focuses on summarizing prototype-based few-shot medical image semantic segmentation algorithms. We analyze network architectures from a problem-oriented perspective and critically examine limitations in current methodologies. Finally, grounded in our analysis of current challenges, we propose forward-looking perspectives for future advancements.