Decoupling mean teacher via Dual students: a prototype-based framework for semi-supervised medical segmentation

Abstract In recent years, Semi-Supervised Learning (SSL) has emerged as a promising approach for medical image segmentation, particularly in scenarios with limited labeled data. Among the existing SSL methods, the mean teacher framework is one of the most popular approaches due to its simple archite...

Full description

Saved in:
Bibliographic Details
Main Authors: Yang Zuo, Renfeng Zhang, Xiurui Guo, Chunmeng Kang, Lei Lyu
Format: Article
Language:English
Published: Springer 2025-06-01
Series:Complex & Intelligent Systems
Subjects:
Online Access:https://doi.org/10.1007/s40747-025-01964-z
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Abstract In recent years, Semi-Supervised Learning (SSL) has emerged as a promising approach for medical image segmentation, particularly in scenarios with limited labeled data. Among the existing SSL methods, the mean teacher framework is one of the most popular approaches due to its simple architecture and outstanding performance. However, this framework has two critical challenges: (1) parameter coupling between the teacher and student models, and (2) unreliable pseudo label generation. To address these limitations, we propose a novel Dual-Student Mean Teacher (DS-MT) framework with three key innovations. First, we advocate the use of two student models rather than the conventional single-teacher-single-student architecture, enabling parameter decoupling through alternative updates between the student models. Second, we develop a prototype-based pseudo-label denoising scheme that evaluates label reliability by calculating the distance between voxel features and prototypes. Last, we implement the Mutual Cross Supervision (MCS) mechanism that facilitates knowledge sharing between the two student models, enabling them to utilize each other’s high-quality predictions as supervised signals. Extensive experiments on three challenging datasets demonstrate that our method outperforms existing state-of-the-art approaches. More importantly, our method is robust across different anatomical structures and imaging modalities, validating its effectiveness in label-limited clinical scenarios.
ISSN:2199-4536
2198-6053