Assessing the Performance of the DINOv2 Self-supervised Learning Vision Transformer Model for the Segmentation of the...

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine
• Cardiovascular Imaging

Background:

• Accurate left atrium segmentation is crucial for diagnosing and treating atrial fibrillation.
• Deep learning models for segmentation typically require large annotated datasets.
• Foundation models offer potential to reduce data dependency through transfer learning.

Purpose of the Study:

• To evaluate the out-of-the-box performance of the DINOv2 foundation model for left atrium segmentation using MRI images.
• To assess DINOv2's adaptability and generalization capabilities with limited annotated data.
• To demonstrate DINOv2's potential for accurate and consistent cardiac segmentation.

Main Methods:

• Utilized DINOv2, a self-supervised vision transformer-based foundation model, for left atrium segmentation.
• Performed end-to-end fine-tuning on MRI datasets.
• Conducted data-level few-shot learning experiments across various dataset sizes.

Main Results:

• Achieved a mean Dice score of 87.1% and an Intersection over Union (IoU) of 79.2% for left atrium segmentation.
• DINOv2 consistently outperformed baseline models across different dataset sizes and patient counts.
• Demonstrated effective adaptation and generalization to MRI data with minimal fine-tuning.

Conclusions:

• DINOv2 shows significant potential as a competitive tool for cardiac segmentation, even with limited data.
• The model provides accurate results essential for pre-procedural planning and pre-operative applications.
• Highlights DINOv2's utility for broader implementation in medical imaging modalities.