Original title: CMRVision: A Foundation Model for Cardiac MR Image Analysis
Athira J. Jacob, Puneet Sharma, Daniel Rueckert · September 2026
Doctors use cardiac magnetic resonance imaging, or CMR, to see heart anatomy, function, and tissue. CMR creates many views and sequence types, which makes building an artificial intelligence model to read all of them difficult. To solve this, researchers built CMRVision. They used DINOv3-styleA specific technique for self-supervised learning that helps the model learn visual features. self-supervised learningA method where AI learns from unlabeled data by finding patterns on its own., a method where an algorithm trains itself to spot patterns and features without human labels, on 36 million CMR images collected from multiple centers. The researchers discovered that training directly on heart scans, using smaller image patches, and adding patch-level goals boosted results. CMRVision beat older models on identifying heart structures, called segmentation. It outlined parts like the left ventricle, right ventricle, and myocardium across different scans. It also showed cross-view generalization by achieving an average Dice score of 0.692 on unseen views, and it reached the top average accuracy of 0.906 when identifying cine views.
When you learn to recognize a friend's face, you usually see them in familiar places like school or home. If you only know them from one setting, you might not spot them at a costume party with different lighting. In this paper, the model is trained on many types of heart scans so it can identify heart parts even when the scan looks different, like from a new angle or with different contrast.
Cardiologists spend hours analyzing heart scans that vary in type and angle, slowing diagnoses. This model could speed up and standardize the process, helping patients get answers faster.
The abstract notes that improvements over prior baselines were modest, though consistent. The model is evaluated only on cardiac MRI tasks, so its generalizability to other medical images is not established. The zero-shot segmentation performance, while demonstrating cross-view generalization, achieved an average Dice score of 0.692, which may not be sufficient for all clinical applications.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
CMRVision: A Foundation Model for Cardiac MR Image Analysis — Athira J. Jacob, Puneet Sharma, Daniel Rueckert
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.