Related Experiment Video
Updated: Jul 12, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
Light mixed-supervised segmentation for 3D medical image data.
Hongxu Yang1, Tao Tan1, Pal Tegzes2
1GE Healthcare, Eindhoven, The Netherlands.
Medical Physics
|November 1, 2023
Summary
This study introduces a mixed-supervised learning method for 3D medical image segmentation, significantly reducing annotation effort. The approach achieves stable and accurate segmentation even with relaxed bounding box annotations, outperforming state-of-the-art methods.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Accurate 3D semantic segmentation is crucial for clinical applications.
- Voxel-level annotation for 3D medical data is labor-intensive and raises privacy concerns.
- Current slice-by-slice annotation methods are time-consuming.
Purpose of the Study:
- To develop a 3D segmentation model that reduces annotation effort.
- To overcome the limitations of existing weakly supervised methods that require tight bounding boxes.
- To enable stable model training using relaxed bounding box annotations.
Main Methods:
- A mixed-supervised training strategy is proposed for 3D segmentation.
- Only one slice requires full contour annotation; others use relaxed bounding boxes.
- The method integrates fully supervised learning, relaxed bounding box priors, and contrastive learning.
Main Results:
- Achieved high segmentation Dice scores: 85.3% on MRI prostate and 83.3% on Vestibular Schwannoma datasets.
- Outperformed state-of-the-art methods using relaxed bounding box annotations.
- Demonstrated stable model performance despite variations in bounding box accuracy.
Conclusions:
- A mixed-supervised learning method for 3D medical imaging is presented.
- The approach allows for stable segmentation with reduced annotation accuracy requirements.
- Facilitates easier model training on large-scale medical datasets.
