UHR-Net: An Uncertainty-Aware Hypergraph Refinement Network for Medical Image Segmentation

Problem
This paper addresses the challenge of accurate lesion segmentation in medical imaging, particularly in scenarios where lesions closely resemble surrounding tissues and have poorly defined boundaries. These characteristics lead to unstable predictions, especially in boundary and transition regions. Additionally, the authors highlight the issue of small-lesion cues being diluted during multi-scale feature extraction, which can result in under- or over-segmentation. The work is presented as a preprint, indicating it has not yet undergone peer review.

Method
The authors propose the Uncertainty-Aware Hypergraph Refinement Network (UHR-Net), which incorporates two key innovations. First, the Uncertainty-Oriented Instance Contrastive (UO-IC) pretraining strategy is introduced. This strategy employs geometry-aware copy-paste augmentation combined with hard-negative mining from lesion-like background regions to enhance instance-level discrimination, particularly for small and visually ambiguous lesions. Second, the architecture features an Uncertainty-Guided Hypergraph Refinement (UGHR) block. This block generates an entropy-based uncertainty map from a coarse probability map, which is then used to guide hypergraph refinement. By categorizing hyperedge prototypes into foreground and background groups, UGHR effectively decouples higher-order interactions, thereby improving segmentation accuracy in ambiguous regions. The training compute details are not disclosed.

Results
UHR-Net was evaluated on five public benchmarks, demonstrating consistent performance improvements over strong baselines. While specific headline numbers are not provided in the abstract, the authors claim that their method outperforms existing techniques, particularly in challenging segmentation tasks involving small and ambiguous lesions. The results suggest significant effect sizes, although exact metrics (e.g., Dice coefficients, Intersection over Union) are not detailed in the summary.

Limitations
The authors acknowledge that their approach may still struggle with extreme cases of ambiguity where lesions are indistinguishable from surrounding tissues. Additionally, the reliance on pretraining may limit the model’s adaptability to datasets with different characteristics. The paper does not discuss potential computational overhead introduced by the hypergraph refinement process or the scalability of the method to larger datasets.

Why it matters
The implications of this work are significant for the field of medical image segmentation, particularly in enhancing the reliability of lesion detection and characterization. By addressing the challenges of ambiguity and small-lesion representation, UHR-Net could improve clinical outcomes through more accurate segmentation, thereby aiding in diagnosis and treatment planning. The proposed methods may also inspire further research into uncertainty-aware techniques and hypergraph-based approaches in other domains of medical imaging and beyond.

Authors: Shuokun Cheng, Jinghao Shi, Kun Sun
Source: arXiv:2604.28095
URL: https://arxiv.org/abs/2604.28095v1