H2AL: Hyperbolic Hierarchy-Aware Learning for Few-Shot Medical Image Segmentation

A novel approach for registration-based few-shot medical image segmentation (RFMIS) has been introduced, tackling the shortcomings of current techniques that function within Euclidean space and consider anatomical structures as flat and separate. This new framework, called H2AL (Hyperbolic Hierarchy-aware Aggregative Learning), improves both the plausibility of deformations and the discrimination of anatomical features for dual-task learning. It features a Hyperbolic Hierarchy-aware Infusion (H2I) module, which utilizes the hierarchical modeling strengths of hyperbolic space to create accurate hierarchy-aware representations through transformation-guided supervised hyperbolic contrastive learning. Detailed in a paper on arXiv (arXiv:2608.07340) marked as cross, this research seeks to enhance the quality of pseudo-labels and segmentation outcomes in medical imaging, especially when labeled data is scarce.