Summary
HeartLens is a computer vision AI-powered tool designed to detect coronary artery calcification (CAC) in CT scans. This project leverages DINOv2, a self-supervised Vision Transformer model, and a U-Net architecture, standard in medical image segmentation. DINO (Self-Distillation with No Labels), developed by Facebook AI Research, enables Vision Transformer to learn semantically rich visual features without manual annotations. In this case of medical images, labeled datasets are limited and timely to produce, which is why DINO is a compelling alternative to supervised pretraining.
A pre-print is available with more details on the novel DINO training process: [2411.07976v7] DINO-LG: A Task-Specific DINO Model for Coronary Calcium Scoring
Model & Data
Our approach utilizes a self-supervised learning framework based on the DINOv2 model, a Vision Transformer architecture designed for computer vision tasks. Our local, secured instance of the DINOv2 model was trained on a dataset of public and internal CT imaging data. Public CT imaging datasets from Stanford were used in the training and testing process, as well as internal, University of Kentucky, CT data extracted by CCTS.
To generate feature maps for calcified areas, we employ a register-based approach. Register maps are feature maps generated by our specialized DINOv2 model. We apply PCA analysis to these four register maps to produce a feature map that highlights and precisely locates calcified areas. This specialized DINOv2 model with registers is fully compatible with CT scans and has been trained using our proprietary Label-Guided data augmentation technique. Additionally, the model is designed to work with single-channel inputs, which is appropriate since most medical images are represented as single-channel images.
The resulting model is capable of detecting coronary artery calcification (CAC) in CT scans with high accuracy, making it a valuable tool for clinicians and researchers in the field of cardiovascular medicine.
Results
HeartLens is still in development. Next steps include clinical evaluation, where our partners will assess the impact of AI-assisted CAC detection on patient care and decision-making.