Related Experiment Video
Updated: Sep 6, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Segmentation only uses sparse annotations: Unified weakly and semi-supervised learning in medical images
Feng Gao1, Minhao Hu2, Min-Er Zhong1
1Department of Colorectal Surgery, The Sixth Affiliated Hospital, Sun Yat-sen University, Guangzhou, Guangdong Province 510655, China; Guangdong Provincial Key Laboratory of Colorectal and Pelvic Floor Diseases, The Sixth Affiliated Hospital, Sun Yat-sen University, Guangzhou, Guangdong Province 510655, China.
This study introduces a new framework for medical image segmentation that learns from limited annotations and abundant unlabeled data. The method achieves state-of-the-art performance, rivaling fully supervised approaches with less data.
Area of Science:
- Medical Image Analysis
- Computer Vision
- Machine Learning
Background:
- Medical image segmentation is crucial but hindered by the time-consuming and expertise-intensive nature of manual annotation.
- Acquiring large, high-quality annotated datasets for segmentation is a significant bottleneck in medical AI development.
Purpose of the Study:
- To develop a novel weakly- and semi-supervised framework for medical image segmentation.
- To enable learning from minimal sparse annotations and large volumes of unlabeled data, reducing annotation burden.
Main Methods:
- Proposed SOUSA (Segmentation Only Uses Sparse Annotations) framework utilizing a teacher-student model architecture.
- Weak supervision of the student model via scribbles and derived Geodesic distance maps.
- Consistency regularization between teacher and student predictions on perturbed unlabeled data using MSE and MPR losses.
Main Results:
- Demonstrated robustness and generalization capabilities across multiple datasets.
- Outperformed existing weakly- and semi-supervised state-of-the-art methods in medical image segmentation.
- Achieved competitive performance compared to fully supervised methods with dense annotations, especially in limited data scenarios.
Conclusions:
- The SOUSA framework effectively reduces the reliance on extensive manual annotations for medical image segmentation.
- This approach offers a viable solution for developing high-performance segmentation models with significantly less labeled data.
- The method shows strong potential for practical application in clinical settings where annotated data is scarce.
