Related Experiment Video
Updated: Jul 29, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
Annotation-efficient training of medical image segmentation network based on scribble guidance in difficult areas.
Mingrui Zhuang1, Zhonghua Chen1,2, Yuxin Yang1
1School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology, Dalian, 116024, China.
Summary
This study introduces an efficient method for training deep medical image segmentation networks using minimal annotations. Focusing human supervision on difficult areas significantly reduces annotation time while maintaining segmentation accuracy for complex cases.
Area of Science:
- Medical Image Analysis
- Deep Learning
- Computer-Aided Diagnosis
Background:
- Deep medical image segmentation requires extensive human-annotated data.
- Existing semi- or non-supervised methods struggle with complex clinical scenarios, leading to inaccurate segmentation in challenging regions like heterogeneous tumors and fuzzy boundaries.
Purpose of the Study:
- To develop an annotation-efficient training approach for deep medical image segmentation networks.
- To reduce the burden of human annotation while improving segmentation accuracy in difficult areas.
Main Methods:
- Proposed an approach using scribble guidance in difficult areas for annotation-efficient training.
- Initially trained a segmentation network with limited fully annotated data, then generated pseudo-labels.
- Utilized probability-modulated geodesic transform for converting scribbles to pseudo-label maps.
- Generated confidence maps by considering geodesic distance and network output probability to mitigate pseudo-label errors.
- Iteratively optimized pseudo-labels and confidence maps with network updates.
Main Results:
- Demonstrated significant reduction in annotation time across brain tumor MRI and liver tumor CT datasets.
- Achieved comparable segmentation accuracy to fully annotated methods with substantially less annotation effort (90 scribble-annotated images vs. >100h for full annotation).
- Maintained high segmentation accuracy in difficult areas, including tumors.
Conclusions:
- The proposed method significantly reduces annotation efforts by concentrating human supervision on challenging regions.
- Offers an efficient solution for training medical image segmentation networks in complex clinical settings.
- Provides a practical alternative to conventional full annotation methods.
