75.64%
Highest average Dice score achieved across 11 organ and tumor categories on the MOTS benchmark.
Research & Publications
CVPR 2021 · IEEE/CVF Conference on Computer Vision and Pattern Recognition
DoDNet: Dynamic on-demand segmentation for multi-organ and tumor analysis
Jianpeng Zhang, Yutong Xie, Yong Xia, Chunhua Shen
2021
Jianpeng Zhang, Yutong Xie, Yong Xia, Chunhua Shen
2021
Publication summary
DoDNet addresses a major challenge in medical image segmentation: learning from multiple datasets where each dataset contains annotations for only specific organs or tumors. Traditional approaches require either multiple independent networks or multi-head architectures, increasing complexity and computational cost.
The proposed Dynamic on-Demand Network (DoDNet) introduces a shared encoder-decoder combined with task encoding, dynamic filter generation, and a dynamic segmentation head. Instead of maintaining dedicated decoders for every task, DoDNet dynamically generates task-specific filters conditioned on both the input image and segmentation task.
Experiments on the newly created MOTS dataset, which combines seven organ and tumor segmentation tasks across 1,155 CT scans, demonstrate that DoDNet achieves superior segmentation performance while maintaining high efficiency. The learned representation also transfers effectively to downstream segmentation tasks.
“DoDNet demonstrates how dynamic task-aware architectures can unify multiple partially labeled medical datasets into a single efficient segmentation framework without sacrificing accuracy.”
Editorial research summary
OrthoAI research content adaptation
Key findings
19.50
Best overall average Hausdorff Distance among competing partially labeled segmentation methods.
920
Large-scale MOTS training dataset enabled transferable pretraining and improved downstream segmentation performance.
Method
Task-aware dynamic segmentation
DoDNet treats partially labeled segmentation as multiple task-specific binary segmentation problems. A task encoding module converts the target organ or tumor into a one-hot representation that informs the network which structure should be segmented.
This task representation is combined with image features and supplied to a controller that dynamically generates segmentation filters tailored to the requested task.
Dynamic filter generation and segmentation head
Unlike conventional segmentation networks where kernels remain fixed after training, DoDNet dynamically generates task-specific convolution kernels for every prediction. The controller conditions filter generation on both image content and task identity.
The lightweight dynamic head enables efficient inference while maintaining flexibility across liver, kidney, pancreas, spleen, vessel, and tumor segmentation tasks.
Key contributions
DoDNet introduces a unified framework for learning from multiple partially labeled medical image datasets.
- Proposes a single dynamic segmentation network capable of handling multiple organ and tumor segmentation tasks.
- Introduces task encoding to explicitly represent segmentation objectives.
- Generates task-specific segmentation filters dynamically using image features and task priors.
- Eliminates the need for multiple dedicated networks or large multi-head architectures.
- Creates the MOTS dataset containing seven partially labeled organ and tumor segmentation tasks.
- Demonstrates effective transfer learning from MOTS pretraining to downstream abdominal segmentation benchmarks.
Similar publications
view all
CVPR 2023 · IEEE/CVF Conference on Computer Vision and Pattern Recognition
PEFAT: Boosting Semi-Supervised Medical Image Classification via Pseudo-loss Estimation and Feature Adversarial Training
Qingjie Zeng, Yutong Xie, Zilin Lu, Yong Xia
CVPR 2024 · IEEE/CVF Conference on Computer Vision and Pattern Recognition
Continual Self-supervised Learning: Towards Universal Multi-modal Medical Data Representation Learning
Yiwen Ye, Yutong Xie, Jianpeng Zhang, Ziyang Chen, Qi Wu, Yong Xia