Multi-Prototype-based Embedding Refinement for Medical Image Segmentation

毕雅莉¹ 车恩羽¹ 陈怡楠¹ 何垣鹏² 瞿经纬^1∗
¹西南大学 ²北京大学
[PDF] [Code]

Abstract

Medical image segmentation aims to identify anatomical structures at the voxel-level. Segmentation accuracy relies on distinguishing voxel differences. Compared to advancements achieved in studies of the inter-class variance, the intra-class variance receives less attention. Moreover, traditional linear classifiers, limited by a single learnable weight per class, struggle to capture this finer distinction. To address the above challenges, we propose a Multi-Prototype-based Embedding Refinement method for semi-supervised medical image segmentation. Specifically, we design a multi-prototype-based classification strategy, rethinking the segmentation from the perspective of structural relationships between voxel embeddings. The intra-class variations are explored by clustering voxels along the distribution of multiple prototypes in each class. Next, we introduce a consistency constraint to alleviate the limitation of linear classifiers. This constraint integrates different classification granularities from a linear classifier and the proposed prototype-based classifier. In the thorough evaluation on two popular benchmarks, our method achieves superior performance compared with state-of-the-art methods.

Multi-Prototype-based Classification

Architecture of MPER

Quantitative Results

Comparison of segmentation quality on the LA dataset. "Labeled" and "Unlabeled" indicate the number and proportion of labeled and unlabeled samples, respectively. Values of four metrics, Dice coefficient, Jaccard index, 95HD, and ASD, are reported. Numbers in bold indicate the best performance.

Comparison of segmentation quality on the ACDC dataset.

Qualitative Results

3D segmentation visualization on the LA dataset.

2D segmentation visualization on the ACDC dataset.

Reference


								@inproceedings{bi2025multi, 

									        title={Multi-Prototype-based Embedding Refinement for Medical Image Segmentation},

									        author={Bi, Yali and Che, Enyu and Chen, Yinan and He, Yuanpeng and Qu, Jingwei},

									        booktitle={Proceedings of the IEEE International Conference on Acoustics, Speech and Signal
									        Processing},

									        pages={1--5},

									        year={2025},

									        publisher={IEEE},

									        address={New York}

									        doi={10.1109/ICASSP49660.2025.10889249}

								}