Mixed Prototype Consistency Learning for Semi-supervised Medical Image Segmentation

TL;DR

This work tackles label scarcity in semi-supervised medical image segmentation by introducing Mixed Prototype Consistency Learning (MPCL), which jointly uses a Mean Teacher and an auxiliary network to generate and fuse mixed prototypes. The framework creates labeled, unlabeled, and mixed prototypes ($p^c_l$, $p^c_u$, $p^c_m$) and fuses them to form a high-quality global prototype $p^c$, guided by temporal ensembling with a time-varying coefficient $\lambda_{con}$ and cosine similarity-based consistency losses. Key contributions include the auxiliary network for mixed data, comprehensive prototype fusion ( $p^c_{lm}$ and $p^c_{um}$ ), and extensive ablations showing robust improvements on Left Atrium and TBAD datasets, especially with limited labeled data. MPCL demonstrates strong practical impact for clinical-grade segmentation under annotation scarcity and offers a scalable approach to leverage mixed data for richer semantic representations.

Abstract

Recently, prototype learning has emerged in semi-supervised medical image segmentation and achieved remarkable performance. However, the scarcity of labeled data limits the expressiveness of prototypes in previous methods, potentially hindering the complete representation of prototypes for class embedding. To address this problem, we propose the Mixed Prototype Consistency Learning (MPCL) framework, which includes a Mean Teacher and an auxiliary network. The Mean Teacher generates prototypes for labeled and unlabeled data, while the auxiliary network produces additional prototypes for mixed data processed by CutMix. Through prototype fusion, mixed prototypes provide extra semantic information to both labeled and unlabeled prototypes. High-quality global prototypes for each class are formed by fusing two enhanced prototypes, optimizing the distribution of hidden embeddings used in consistency learning. Extensive experiments on the left atrium and type B aortic dissection datasets demonstrate MPCL's superiority over previous state-of-the-art approaches, confirming the effectiveness of our framework. The code will be released soon.

Figures (2)

• Figure 1: The figure illustrates the structure of the proposed Mixed Prototype Consistency Learning framework, which is composed of a Mean Teacher structure and an auxiliary network. The student and teacher networks are utilized to generate labeled and unlabeled prototypes $p^c_l, p^c_u$, respectively. And the auxiliary network produces mixed prototypes $p^c_m$ for mixed data processed by CutMix. The mixed prototype fusion module includes three fusion processes. The labeled and unlabeled prototypes will be fused with mixed prototypes to enhance their semantic information. Finally, a high-quality global prototype representation $p^c$ is formed by fusing labeled and unlabeled prototypes, which optimizes the distribution of hidden embeddings in consistency learning.
• Figure 2: The visualizations of experimental results on Left Atrium and Aortic Dissection datasets. MPCL denotes the proposed mixed prototypes consistency learning. GT denotes the ground truth labels. Baseline represents the method of URPC. ITK-SNAP py06nimg is the tool to visualize 3D medical image.