Cardiac MR foundation model pretrained DINOv3-style on 36 million images, fine-tuned to segment cine, LGE and mapping scans and classify cine views.
No providers recorded yet. Browse all providers
The myocardium on a cardiac magnetic resonance (CMR) scan is a ring a few millimetres thick, and it looks different in every sequence a clinician acquires: dark against bright blood on cine, patchily bright where scar takes up gadolinium on late gadolinium enhancement (LGE), a smooth field of relaxation values on T1 and T2 maps. A vision transformer that chops a 224×224 image into 16×16-pixel patches spends only a couple of tokens across that wall, and features learned from natural photographs have never seen an inversion-recovery contrast at all. Both problems bite hardest on exactly the sequences where annotated data is scarcest.
CMRVision addresses them with a CMR-specific encoder trained on the full range of sequences. It adapts the DINOv3 self-supervised recipe to cardiac MR: a ViT-S student and teacher are trained to agree on global views (a DINO loss), on individual patch embeddings (an iBOT loss), and on the pairwise correlation structure among patches (Gram anchoring), with KoLeo regularization spreading the embeddings out. The patch size is halved to 8×8 so the thin myocardial wall and small atrial structures occupy more tokens, and augmentations that make no sense for MR, such as color jitter and vertical flips, are removed.
It comes from Siemens Healthineers with the Technical University of Munich and Imperial College London, the group behind CMR-SSL. It reuses that model's 36-million-image pretraining cohort but replaces its DINO v1 objective with the DINOv3-style one, and it benchmarks CMR-SSL directly as a baseline. The preprint appeared on arXiv on 1 September 2026 and was accepted at the MedAGI 2026 workshop at MICCAI 2026.
Pretraining used over 36 million unlabeled images from about 27,000 studies: 16,230 UK Biobank participants (short-axis and long-axis cine), 10,073 patients from one clinical center (cine, LGE, T2-weighted, perfusion, localizers) and 1,221 from a second (including LV-outflow and aortic cines), on 1.5 T and 3 T Siemens scanners. Images were resampled to 1 mm in-plane, resized to 224×224 and trained in two stages, half then full resolution (batch sizes 1024 and 256), with 2 global and 6 local crops, axial rotary position embeddings, and initialization from natural-image weights, for seven days using PyTorch Lightning and LightlySSL.
Downstream, the encoder is fine-tuned end to end inside a UNETR for segmentation and with a linear head for classification. On a benchmark of 1,873 patients and 11,022 images drawn from ACDC, a Kaggle long-axis cohort, EMIDEC and internal centers, it scored Dice 0.940–0.967 for the left ventricle, 0.855–0.905 for myocardium, 0.929 for the right ventricle, 0.920 for the left atrium and 0.931 for the right atrium, best or tied on all 11 targets against SAM2, MedSAM2, UNETR, SwinUNETR, CineMA and CMR-SSL. Margins over CMR-SSL are small, largest for LGE myocardium (0.888 against 0.876). Zero-shot LGE long-axis myocardium Dice was 0.692 (0.794 on 4-chamber, 0.612 on 2-chamber). Cine view classification reached 0.906 average accuracy against 0.894 for CMR-SSL, while a DINOv2-style variant trained the same way scored 0.908.
The target is routine CMR post-processing: contouring chambers and myocardium for volumes, ejection fraction and mass, delineating the myocardium on LGE and T1/T2 maps for tissue characterization, and sorting cine series by view. For teams building CMR tools, the encoder is also the starting point for other models. MR-JEPA, a video foundation model from the same group, initializes its spatiotemporal encoder from a 2D CMR model trained with the DINOv3 objective.
CMRVision makes a design argument more than a leaderboard one: for dense cardiac prediction, patch-level self-supervision and small patches matter, and domain pretraining beats adapting natural-image features. Its gains over CMR-SSL are modest, and view classification barely separates CMR-pretrained variants. The model is 2D only, with no volumetric or temporal modelling, and some downstream patients from the two internal centers may overlap the pretraining corpus. No code or weights have been released, and the paper describes it as research not commercially available, so outside groups cannot use it directly. Its reach is indirect: as the base for MR-JEPA and as evidence for how to build the next CMR encoder.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.