Optimizing retinal images based carotid atherosclerosis prediction with explainable foundation models.

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Cardiovascular Disease Research

Background:

• Carotid atherosclerosis is a significant predictor of cardiovascular disease (CVD).
• Early detection of atherosclerosis is crucial for CVD prevention.
• The optimal selection and fine-tuning of foundation models (FMs) for classifying carotid atherosclerosis from retinal images require investigation.

Purpose of the Study:

• To evaluate the performance of different vision foundation models (FMs) and fine-tuning strategies for detecting carotid atherosclerosis from retinal images.
• To assess the clinical utility and prognostic relevance of these models in predicting future CVD mortality.
• To determine the explainability and vascular alignment of the best-performing models.

Main Methods:

• Utilized a dataset of 39,620 individuals.
• Evaluated four vision FMs with three distinct fine-tuning methods.
• Assessed performance using predictive metrics (AUC, sensitivity, specificity), survival analysis for CVD mortality, and Grad-CAM with vessel segmentation for explainability.

Main Results:

• DINOv2 with low-rank adaptation achieved the highest performance (AUC=0.71, sensitivity=0.87, specificity=0.44).
• The model demonstrated significant prognostic relevance for future CVD mortality (HR=2.20, P-trend<0.05).
• The selected model exhibited strong vascular alignment, indicating good explainability.

Conclusions:

• Foundation models, particularly DINOv2, show promise for classifying carotid atherosclerosis from retinal images.
• The findings support the feasibility of opportunistic screening for atherosclerosis and CVD using retinal imaging.
• A multi-dimensional evaluation framework is essential for selecting optimal FMs in medical AI.