Self-supervised Vision Transformer for coronary artery calcium (CAC) detection. Uses DINOv2 with register-based approach — PCA on 4 register maps produces calcification feature maps. Employs Label-Guided data augmentation technique.
Works on both gated and non-gated CT scans. Color-codes severity for clinical review. Supports classification, segmentation, and LLM-assisted diagnostic reporting.
Long-term goal: EHR integration for point-of-care clinical decision-making. Potential to detect stenosis and other conditions.
Training data: Public Stanford CT datasets + internal CT data. Single-channel input designed for medical imaging.
HeartLens is still in development. Contact ai@uky.edu for collaboration.
This model is hosted on Hugging Face.