Partial Label Multi-organ Segmentation based on Local Feature Enhancement.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Automatic segmentation of abdominal organs from CT images is crucial for surgical planning.
• Partial annotation of organs across different medical institutions presents a significant challenge for multi-center studies.

Purpose of the Study:

• To develop a robust 3D local feature enhanced multi-head segmentation network.
• To address the partial annotation problem in multi-center abdominal multi-organ segmentation.

Main Methods:

• A novel architecture combining a global branch (3D Transformer and U-Net fusion - 3D TransUNet) and a local 3D U-Net branch.
• The local branch enhances the global branch with additional abdominal organ structure information.
• Evaluation on four public CT datasets with varying partial labels.

Main Results:

• Achieved an average Dice Score Coefficient (DSC) of 93.01%.
• Obtained an average Hausdorff Distance (HD) of 3.489 mm.
• Demonstrated superior accuracy and robustness compared to three state-of-the-art methods.

Conclusions:

• The proposed 3D local feature enhanced multi-head segmentation network effectively handles partially labeled datasets.
• This method offers improved accuracy and robustness for multi-center abdominal multi-organ segmentation.
• The approach shows significant potential for enhancing surgical planning in abdominal diseases.