Related Experiment Video
Updated: Jan 10, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
DMformer: Difficulty-Adapted Masked Transformer for Semi-Supervised Medical Image Segmentation.
IEEE Journal of Biomedical and Health Informatics
|November 26, 2025
Summary
This study introduces a novel Difficulty-adapted Masked Transformer (DMformer) for semi-supervised medical image segmentation. DMformer enhances learning by adapting reconstruction difficulty, significantly improving segmentation accuracy with limited labeled data.
Area of Science:
- Medical Image Analysis
- Computer Vision
- Machine Learning
Background:
- Leveraging unlabeled data is crucial for semi-supervised medical image segmentation.
- Shared human anatomy provides a strong prior for utilizing unlabeled medical images.
- Masked image modeling inspires new approaches for incorporating anatomical priors.
Purpose of the Study:
- To develop a semi-supervised medical image segmentation framework that effectively utilizes unlabeled data through anatomical priors.
- To introduce a difficulty-adapted mask mechanism to handle varying reconstruction complexities of different organs/tissues.
- To improve the performance of medical image segmentation models using limited labeled data.
Main Methods:
- Incorporated an auxiliary unsupervised gross anatomy reconstruction task into a teacher-student framework.
- Developed a difficulty-adapted mask mechanism by modulating masked region and class ratios.
- Implemented region-based and class-based masking strategies tailored to reconstruction difficulty.
- Utilized a conflict-aware gradient computation strategy to manage simultaneous factor modulation.
- Built the Difficulty-adapted Masked Transformer (DMformer) upon vision transformers.
Main Results:
- DMformer demonstrated superior performance in semi-supervised medical image segmentation.
- Achieved significant improvements in Dice Similarity Coefficient (DSC) on ACDC and Synapse datasets.
- Outperformed state-of-the-art (SOTA) methods with 5% labeled images on ACDC (9.53% DSC improvement).
- Outperformed SOTA methods with 30% labeled images on Synapse (4.63% DSC improvement).
Conclusions:
- The proposed difficulty-adapted mask mechanism effectively addresses varying reconstruction challenges in medical image segmentation.
- DMformer offers a powerful approach for semi-supervised medical image segmentation, especially with limited labeled data.
- The framework successfully leverages anatomical priors through an auxiliary reconstruction task, enhancing segmentation accuracy.
