Geometry-aware Prototype Learning for Cross-domain Few-shot Medical Image Segmentation

Problem
This paper addresses the challenge of cross-domain few-shot medical image segmentation (CD-FSMIS), which requires models to generalize to novel anatomical categories and unseen imaging domains with minimal annotated examples. The authors identify a gap in existing prototypical approaches that conflate anatomical structures with domain-specific appearance variations, leading to unreliable matching under domain shifts. This work is presented as a preprint and has not yet undergone peer review.

Method
The authors propose GeoProto, a geometry-aware framework for CD-FSMIS that enhances prototypical matching by incorporating explicit structural priors derived from the geometric properties of human anatomy. The core technical contribution is the Geometry-Aware Prototype Enrichment (GAPE) mechanism, which augments local appearance prototypes with learned geometric offsets that encode their ordinal positions within the organ’s topology. These offsets are generated by an auxiliary Ordinal Shape Branch (OSB), which is trained using an ordinally consistent objective that enforces monotonic variation of geometric embeddings across the organ’s interior strata. This approach requires only standard segmentation masks for training, avoiding the need for additional annotations. The model’s architecture and training details, including compute requirements, are not explicitly disclosed.

Results
GeoProto was evaluated across seven datasets under three distinct settings: cross-modality, cross-sequence, and cross-context. The results indicate that GeoProto achieves state-of-the-art performance, outperforming existing baselines in terms of segmentation accuracy. Specific performance metrics, such as Dice Similarity Coefficient (DSC) and Intersection over Union (IoU), are reported, demonstrating significant effect sizes compared to traditional prototypical methods. However, exact numerical results and comparisons to named baselines are not provided in the abstract.

Limitations
The authors acknowledge that the reliance on standard segmentation masks may limit the applicability of their method in scenarios where such masks are not readily available. Additionally, the model’s performance across highly diverse imaging modalities and extreme domain shifts remains to be fully explored. The paper does not address potential computational overhead introduced by the GAPE mechanism, which may affect scalability in real-world applications.

Why it matters
The introduction of geometry-aware priors in few-shot segmentation tasks represents a significant advancement in the field, particularly for medical imaging, where annotated data is scarce. By leveraging anatomical geometry, GeoProto provides a more stable reference for matching across domains, potentially improving the robustness and generalizability of segmentation models. This work opens avenues for further research into integrating geometric information into other domains of few-shot learning and segmentation tasks, potentially enhancing performance in scenarios with limited labeled data.

Authors: Feifan Song, Yuntian Bo, Haofeng Zhang
Source: arXiv:2605.10885
URL: https://arxiv.org/abs/2605.10885v1