Cross-Anatomy Transfer Learning via Shape-Aware Adaptive Fine-Tuning for 3D Vessel Segmentation

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Anatomy

Background:

• Deep learning excels at vessel segmentation but requires extensive labeled data.
• Transfer learning can leverage existing datasets but struggles with cross-anatomy domain shift due to structural discrepancies.

Purpose of the Study:

• To develop a robust cross-anatomy transfer learning framework for 3D vessel segmentation.
• To address the performance limitations of transfer learning caused by significant differences in vessel structures across anatomies.

Main Methods:

• A pre-trained model was generated on a public hepatic vessel dataset.
• An adaptive fine-tuning strategy dynamically adjusted network filters for target vessel segmentation.
• A Gaussian-based signed distance map was incorporated as an auxiliary task to capture geometry-aware knowledge.

Main Results:

• The proposed framework effectively overcame the cross-anatomy domain shift.
• Accurate vessel segmentation was achieved on small-scale coronary artery and brain vessel datasets.
• The adaptive fine-tuning and auxiliary task improved model generalization.

Conclusions:

• The developed framework significantly enhances 3D vessel segmentation accuracy across different anatomical sites.
• This approach reduces the need for large, site-specific labeled datasets.
• The method shows promise for broader applications in medical image analysis.