Related Experiment Videos
CED-Diff: a clinically-experience-driven diffusion model with differentiated feature extraction for head and neck
Ye Yuan1, Manyu Cui1, Zhaoshuo Diao1
1School of Software, Shenyang University of Technology, Shenyang, People's Republic of China.
Physics in Medicine and Biology
|July 21, 2026
Summary
We developed CED-Diff, a new AI framework for segmenting head and neck tumors in PET/CT scans. This method improves accuracy for radiotherapy planning by better utilizing complementary PET and CT image data.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiotherapy Planning
Background:
- Accurate segmentation of head and neck tumors from PET/CT is crucial for radiotherapy.
- Existing methods struggle with heterogeneous PET/CT data, impacting small or indistinct lesions.
- Fusion strategies often fail to explicitly model metabolic (PET) and anatomical (CT) information.
Purpose of the Study:
- To introduce CED-Diff, a novel diffusion-based framework for head and neck tumor segmentation.
- To enhance tumor delineation by explicitly integrating complementary PET and CT information.
- To improve the accuracy and robustness of automated segmentation for improved treatment planning.
Main Methods:
- Proposed CED-Diff, a diffusion-based segmentation framework.
- Integrated Differentiated Feature Extraction (DFE) to model metabolic and anatomical data.
- Employed Task-Oriented Auxiliary Supervision (TAS) for improved feature learning.
Main Results:
- CED-Diff achieved the highest Dice score (66.28%) on the HeadNeck dataset.
- Demonstrated superior performance with a Sensitivity of 78.70% and Precision of 69.08%.
- Outperformed existing methods in segmenting head and neck tumors.
Conclusions:
- Explicitly modeling complementary PET and CT information significantly improves segmentation accuracy.
- CED-Diff offers a more robust approach for automatic head and neck tumor segmentation.
- The findings support the clinical utility of advanced AI in radiotherapy workflows.