Towards a general computed tomography image segmentation model for anatomical structures and lesions
Xi Ouyang1, Dongdong Gu1, Xuejian Li1
1Department of Research and Development, United Imaging Intelligence, Shanghai, China.
Communications Engineering
|October 13, 2024
Summary
A new general medical image segmentation model (gCIS) improves computerized tomography (CT) analysis by learning multiple tasks simultaneously. This approach enhances segmentation performance and allows for efficient adaptation to new segmentation challenges.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Deep learning models for medical image segmentation often use task-specific data, overlooking inter-task relationships.
- A unified approach can potentially improve performance and generalization across diverse segmentation tasks.
Purpose of the Study:
- To develop a general medical image segmentation model for computerized tomography (CT) volumes.
- To demonstrate that joint learning across multiple tasks enhances segmentation performance.
- To enable automatic segmentation of various CT tasks using text prompts.
Main Methods:
- Developed the general CT image segmentation (gCIS) model with a shared transformer-based encoder.
- Incorporated automatic pathway modules for task-prompt-based decoding.
- Trained the model on a large dataset of 36,419 CT scans across 83 tasks.
Main Results:
- Achieved an average Dice coefficient of 82.84% across multiple segmentation tasks.
- Demonstrated automatic segmentation of various tasks via text prompts and automatic pathway routing.
- Showcased efficient adaptation to new tasks with minimal training data and parameter pruning capabilities.
Conclusions:
- Jointly learning diverse segmentation tasks with a general model improves CT image segmentation performance.
- The gCIS model offers flexibility, efficiency, and adaptability for medical image analysis.
- Automatic pathway routing enables network optimization and rapid learning for novel tasks.
