Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Boosting Semi-Supervised Medical Image Segmentation via Masked Image Consistency and Discrepancy Learning

Pengcheng Zhou, Lantian Zhang, Wei Li

TL;DR

This work targets semi-supervised medical image segmentation by addressing the trade-off between information exchange and model diversity in co-training. The authors introduce MICD, a heterogeneous dual-branch framework with three modules: MCPC to capture fine-grained local context through masked cross pseudo supervision, CFC to enforce decoder feature consistency, and CMD to maintain branch diversity using EMA teachers. The training objective combines supervised loss with three cross-branch regularizers, employing a Gaussian ramp-up for stabilization. Empirical results on AMOS and Synapse show that MICD surpasses state-of-the-art SSL methods across multiple labeled-data regimes, with notable gains on minority structures, validating the approach’s robustness and practical impact in medical imaging.

Abstract

Semi-supervised learning is of great significance in medical image segmentation by exploiting unlabeled data. Among its strategies, the co-training framework is prominent. However, previous co-training studies predominantly concentrate on network initialization variances and pseudo-label generation, while overlooking the equilibrium between information interchange and model diversity preservation. In this paper, we propose the Masked Image Consistency and Discrepancy Learning (MICD) framework with three key modules. The Masked Cross Pseudo Consistency (MCPC) module enriches context perception and small sample learning via pseudo-labeling across masked-input branches. The Cross Feature Consistency (CFC) module fortifies information exchange and model robustness by ensuring decoder feature consistency. The Cross Model Discrepancy (CMD) module utilizes EMA teacher networks to oversee outputs and preserve branch diversity. Together, these modules address existing limitations by focusing on fine-grained local information and maintaining diversity in a heterogeneous framework. Experiments on two public medical image datasets, AMOS and Synapse, demonstrate that our approach outperforms state-of-the-art methods.

Boosting Semi-Supervised Medical Image Segmentation via Masked Image Consistency and Discrepancy Learning

TL;DR

This work targets semi-supervised medical image segmentation by addressing the trade-off between information exchange and model diversity in co-training. The authors introduce MICD, a heterogeneous dual-branch framework with three modules: MCPC to capture fine-grained local context through masked cross pseudo supervision, CFC to enforce decoder feature consistency, and CMD to maintain branch diversity using EMA teachers. The training objective combines supervised loss with three cross-branch regularizers, employing a Gaussian ramp-up for stabilization. Empirical results on AMOS and Synapse show that MICD surpasses state-of-the-art SSL methods across multiple labeled-data regimes, with notable gains on minority structures, validating the approach’s robustness and practical impact in medical imaging.

Abstract

Semi-supervised learning is of great significance in medical image segmentation by exploiting unlabeled data. Among its strategies, the co-training framework is prominent. However, previous co-training studies predominantly concentrate on network initialization variances and pseudo-label generation, while overlooking the equilibrium between information interchange and model diversity preservation. In this paper, we propose the Masked Image Consistency and Discrepancy Learning (MICD) framework with three key modules. The Masked Cross Pseudo Consistency (MCPC) module enriches context perception and small sample learning via pseudo-labeling across masked-input branches. The Cross Feature Consistency (CFC) module fortifies information exchange and model robustness by ensuring decoder feature consistency. The Cross Model Discrepancy (CMD) module utilizes EMA teacher networks to oversee outputs and preserve branch diversity. Together, these modules address existing limitations by focusing on fine-grained local information and maintaining diversity in a heterogeneous framework. Experiments on two public medical image datasets, AMOS and Synapse, demonstrate that our approach outperforms state-of-the-art methods.

Paper Structure

This paper contains 15 sections, 7 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Overview
  4. Masked Cross Pseudo Consistency
  5. Cross Feature Consistency
  6. Cross Model Discrepancy
  7. Overall Training Objective
  8. Experiments
  9. Experimental Setup
  10. Dataset
  11. Implementation Details
  12. Evaluation metrics
  13. Result and Analysis
  14. Ablation Studies
  15. Conclusion

Figures (2)

  • Figure 1: An overview of the proposed Masked Image Consistency and Discrepancy Learning (MICD) architecture. Our MICD framework consists of two VNet models and their Teacher networks smoothed via EMA, allowing the peer networks to complement each other through various training strategies. These strategies include Masked Cross Pseudo Consistency (MCPC), Cross Feature Consistency (CFC), and Cross Model Discrepancy (CMD).
  • Figure 2: Visual comparison between MICD and the SOTA methods on 5% labeled AMOS dataset. Regions in red boxes show poor segmentation performance, whereas those in green boxes highlight some minority classes being segmented.