Investigating the Domain Adaptability of General-Purpose Foundation Models for Left Atrium Segmentation from MR...

Area of Science:

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

Background:

• Accurate left atrium (LA) segmentation is vital for assessing cardiac health, particularly in diseases like atrial fibrillation (AFib).
• Deep learning models for segmentation typically require large annotated datasets, which are often unavailable for specialized tasks like LA segmentation.
• Pre-trained foundation models offer transferable features, making them suitable for data-scarce medical imaging domains.

Purpose of the Study:

• To investigate the domain adaptability and robustness of foundation models (DINOv2, SAM, MedSAM) for left atrium segmentation from MRI images.
• To evaluate the effectiveness of integrating a UNet decoder with foundation models for improved LA segmentation performance.
• To assess the models' performance in both end-to-end fine-tuning and low-data regimes.

Main Methods:

• Explored DINOv2, SAM, and MedSAM foundation models for LA segmentation using MRI data.
• Integrated a modified UNet decoder to leverage global contextual features from foundation models.
• Evaluated the approach on the 2022 LAScarQS and 2018 LASC segmentation challenge datasets.

Main Results:

• The UNet decoder integration outperformed linear decoders and other UNet baselines.
• The DINOv2 model with a UNet decoder achieved superior Dice scores (91.5%, 91.6%) and IoU scores (84.5%, 86.6%).
• Consistent high performance was observed across diverse datasets and limited training data scenarios, demonstrating generalizability and robustness.

Conclusions:

• Foundation models, particularly DINOv2 combined with a UNet decoder, present a powerful and adaptable solution for left atrium segmentation.
• This approach effectively addresses the challenge of limited annotated data in medical image segmentation.
• The findings highlight the transformative potential of foundation models for developing generalized and robust medical image analysis tools.