StrDiSeg: Adapter-Enhanced DINOv3 for Automated Ischemic Stroke Lesion Segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep vision foundation models (e.g., DINOv3) possess strong visual representation capabilities.
• Direct application in medical image segmentation is hindered by limited annotated data and high computational costs of full fine-tuning.

Purpose of the Study:

• To propose StrDiSeg, an adaptation framework for efficient fine-tuning of deep vision models in medical image segmentation.
• To enable task-specific learning while preserving pretrained knowledge from models like DINOv3.

Main Methods:

• Integration of lightweight bottleneck adapters within transformer layers of DINOv3.
• Utilizing an attention-enhanced U-Net decoder with multi-scale feature fusion.
• Experiments conducted on ischemic stroke lesion segmentation datasets (AISD and ISLES22).

Main Results:

• StrDiSeg achieved Dice scores of 0.516 on AISD (Non-Contrast CT) and 0.824 on ISLES22 (DWI).
• The method outperformed baseline models in segmenting ischemic stroke lesions.
• Demonstrated robustness across different clinical imaging modalities.

Conclusions:

• Adapter-based fine-tuning offers a practical and computationally efficient strategy for medical image segmentation.
• Leveraging large pretrained vision models is feasible for clinical applications with methods like StrDiSeg.
• The framework shows promise for improving segmentation accuracy in challenging medical imaging tasks.