Related Experiment Video
Updated: Jun 29, 2025
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
48.0K
SDMI-Net: Spatially Dependent Mutual Information Network for semi-supervised medical image segmentation.
Di Gai1, Zheng Huang2, Weidong Min1
1School of Mathematics and Computer Science, Nanchang University, Nanchang, 330031, China; Institute of Metaverse, Nanchang University, Nanchang, 330031, China.
Computers in Biology and Medicine
|April 6, 2024
Summary
This study introduces a novel dual-teacher method for semi-supervised medical image segmentation, improving deep learning models using unlabeled data by focusing on block-level consistency and uncertainty estimation for better accuracy.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Semi-supervised learning in medical image segmentation uses limited labeled data with abundant unlabeled data to train deep models.
- Voxel-level consistency learning methods are hindered by low-confidence voxels and ignore spatial correlations between voxels.
- Existing methods struggle with efficiently and accurately segmenting medical images due to these limitations.
Purpose of the Study:
- To develop a robust semi-supervised medical image segmentation method that overcomes the limitations of voxel-level consistency learning.
- To enhance the spatial dependence between neighboring voxels for more reliable segmentation.
- To achieve state-of-the-art performance in medical image segmentation tasks.
Main Methods:
- Proposed a dual-teacher affine consistent uncertainty estimation to filter high-uncertainty voxels, promoting reliable voxel-level learning.
- Introduced a spatially dependent mutual information module to maximize mutual information between local voxel blocks for block-level consistency.
- Implemented and validated the method on the Left Atrial Segmentation Challenge and BraTS-2019 datasets.
Main Results:
- The proposed method effectively filters out uncertain voxels, improving the reliability of the segmentation process.
- Block-level consistency learning significantly enhances the spatial dependence between neighboring voxels.
- Achieved state-of-the-art quantitative and qualitative results on benchmark medical image segmentation datasets.
Conclusions:
- The dual-teacher affine consistent uncertainty estimation and spatially dependent mutual information module offer a superior approach to semi-supervised medical image segmentation.
- This method addresses key limitations of previous techniques, leading to improved segmentation accuracy and robustness.
- The approach demonstrates significant potential for advancing deep learning applications in medical image analysis.
