Related Experiment Video
Updated: Sep 15, 2025
11:38
Volume Segmentation and Analysis of Biological Materials Using SuRVoS Super-region Volume Segmentation Workbench
Published on: August 23, 2017
9.9K
CSCE: Cross Supervising and Confidence Enhancement pseudo-labels for semi-supervised subcortical brain structure
Yuan Sui1, Yujie Zhang1, Chengan Liu1
1School of Computer Science and Engineering, Northeastern University, Shenyang, 110169, Liaoning, China.
Journal of Neuroscience Methods
|July 13, 2025
Summary
This study introduces a new semi-supervised method for segmenting brain structures using dual deep learning models. The approach enhances accuracy by improving pseudo-label reliability, leading to better segmentation results without increased computational cost.
Area of Science:
- Neuroimaging
- Medical Image Analysis
- Artificial Intelligence
Background:
- Accurate segmentation of subcortical brain structures is crucial for diagnosing and treating brain diseases.
- Deep learning models excel at medical image segmentation but require extensive labeled data, which is costly and time-consuming to acquire.
- Semi-supervised learning offers a practical solution by leveraging unlabeled data.
Purpose of the Study:
- To propose a novel semi-supervised framework for subcortical brain structure segmentation using dual student-teacher models.
- To enhance the reliability of pseudo-labels through confidence enhancement mechanisms.
- To achieve improved segmentation performance without increasing computational resources.
Main Methods:
- A dual student-teacher framework employing U-Net and TransUNet models for mutual supervision.
- Pseudo-label generation by teacher models to train student models.
- Confidence enhancement mechanisms including information entropy for uncertainty quantification and an auxiliary detection task for pseudo-label screening.
Main Results:
- The proposed Cross Supervising and Confidence Enhancement (CSCE) framework significantly improved segmentation accuracy on two public brain MRI datasets.
- Achieved superior Dice scores and Minimum Hausdorff Distance (MHD) values compared to state-of-the-art semi-supervised methods.
- The final inference network, using only one teacher model, demonstrated enhanced results without additional parameters or segmentation time.
Conclusions:
- The CSCE framework effectively leverages unlabeled data for robust semi-supervised subcortical brain structure segmentation.
- Confidence enhancement mechanisms are vital for reliable cross-supervision in semi-supervised learning.
- The proposed method offers a computationally efficient and highly accurate solution for clinical applications in neuroimaging.
