Related Experiment Video
Updated: Aug 6, 2026
Automated Joint Space Detection Improves Bone Segmentation Accuracy
Published on: November 28, 2025
Semi-supervised Medical Image Segmentation via Perturbation-Aware Mutual Learning and Edge-Aware Uncertainty Loss for
Waqas Anwaar1,2, Van Manh3, Wufeng Xue4,5
1Guangdong Key Laboratory of Biomedical Measurements and Ultrasound Imaging, Marshall Laboratory of Biomedical Engineering, School of Biomedical Engineering, National-Regional Key Technology Engineering Laboratory for Medical Ultrasound, Shenzhen University Medical School, Shenzhen University, Shenzhen, 518060, China.
This study introduces a new semi-supervised learning framework to improve medical image segmentation, particularly for cardiac structures. The method enhances boundary accuracy by using unlabeled data, outperforming existing techniques in segmentation tasks.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Accurate medical volume segmentation in MRI is crucial for organ structure analysis.
- Supervised learning for segmentation requires extensive expert-labeled data, which is costly and time-consuming.
- Semi-supervised learning (SSL) methods utilize unlabeled data but struggle with edge regions and boundary accuracy.
Purpose of the Study:
- To develop a novel semi-supervised learning framework to enhance medical image segmentation, focusing on cardiac structures.
- To address the limitations of existing SSL methods in accurately segmenting boundary regions.
- To improve the clinical relevance of segmentation by enhancing accuracy in critical areas.
Main Methods:
- A mutual learning module with multiple decoders was designed to generate diverse probability predictions via feature perturbations.
- A dual fine-grained boundary loss and an edge-aware uncertainty loss were implemented for consistency constraints between labeled and unlabeled data.
- The framework was evaluated on 2D (ACDC) and 3D (LA) cardiac datasets, comparing against seven state-of-the-art SSL methods.
Main Results:
- The proposed framework demonstrated superior segmentation performance compared to existing methods on public cardiac datasets.
- Extensive qualitative and quantitative experiments confirmed the effectiveness of the approach in semi-supervised settings.
- Clinically relevant cardiac measurements derived from segmentations showed the framework's practical utility.
Conclusions:
- The novel semi-supervised framework effectively leverages unlabeled data to improve cardiac structure segmentation accuracy, especially at boundaries.
- The dual fine-grained boundary and edge-aware uncertainty losses are key components for enhancing segmentation performance.
- The approach offers a promising solution for accurate and clinically relevant medical image segmentation with reduced reliance on labeled data.