Cine cardiac MRI foundation model: a multi-view masked autoencoder pretrained on 15 million UK Biobank images for segmentation and diagnosis.
No providers recorded yet. Browse all providers
A cine cardiac magnetic resonance (CMR) study is the reference standard for measuring how the heart fills and empties, but turning it into numbers still takes an experienced reader tracing ventricular borders frame by frame, and two readers rarely agree exactly. Deep learning has automated pieces of that work, yet each piece usually arrives as its own network trained from scratch: one for short-axis segmentation, another for long-axis landmarks, a third for disease classification. Each needs its own labelled cohort, and none of them benefits from what the others learned about cardiac anatomy.
CineMA (cine CMR masked autoencoder) replaces that stack with one pretrained backbone. It was developed by Yunguan Fu, Yipeng Hu and colleagues at University College London and InstaDeep, with co-authors at Imperial College London and the Barts Heart Centre of Barts Health NHS Trust. The preprint appeared in May 2025 and the peer-reviewed version in Communications Medicine in 2026. The model learns by hiding three quarters of each image and reconstructing it, so the 74,916 unlabelled UK Biobank studies it trains on teach it cardiac shape and motion without a single annotation. Downstream, the reconstruction decoder is discarded and a small task head is attached.
What distinguishes CineMA from earlier cardiac foundation models such as CMR-SSL is that it treats a CMR exam as the multi-view object it is. The short-axis stack is handled as a 3D volume and the two-, three- and four-chamber long-axis views as 2D images, and all four are encoded jointly by one shared transformer.
CineMA uses the MAE "base" configuration and has about 126 million parameters. Short-axis inputs are resampled to 1 × 1 × 10 mm and cropped to 192 × 192 × 16; long-axis inputs to 1 × 1 mm and 256 × 256. Pretraining masked 75% of 16 × 16 patches and minimized mean squared error on the masked regions. It ran for 800 epochs over UK Biobank data fields 20208 and 20209 (more than 15 million images), with batch size 128 and AdamW, taking 18 days on eight NVIDIA RTX A6000 GPUs.
Evaluation spanned eight downstream datasets: ACDC, M&Ms, M&Ms2, Kaggle, Rescan, a long-axis Landmark set, EMIDEC and MyoPS2020. The same hyperparameters were used throughout, with no per-dataset search. On the pooled ACDC, M&Ms and M&Ms2 test sets, fine-tuned CineMA reached a mean Dice of 88.72% against 89.24% for nnUNet. Applied without retraining to the Kaggle and Rescan cohorts, it estimated left ventricular ejection fraction with 4.65% mean absolute error (nnUNet 4.30%). Across Rescan's repeat-scan pairs its coefficient of variation was 6.73%, against 9.53% for nnUNet. With 10% of the training data those figures became 8.47% and 20.20%. In binary disease detection, specificity rose from 26.8% to 54.9% over a ResNet-50 baseline at higher sensitivity. Landmark localization reached about 1 mm mean error. Gains on LGE and T2-weighted scar segmentation were limited, since those sequences were absent from pretraining.
The released fine-tuned models let a cardiac imaging group segment ventricles and myocardium, estimate ejection fraction, locate mitral-valve and apical landmarks for MAPSE and global longitudinal shortening, and flag likely disease, all without training a network. Groups with their own labels can fine-tune the backbone on new pathologies or cohorts, where it is most useful when labelled data are scarce. At population scale, CineMA-derived measurements and embeddings support biobank studies. In exploratory UK Biobank analyses the encoder predicted prevalent diabetes (AUROC 0.871) and hypertension (0.771), and three-year survival (0.748).
CineMA brought large-scale masked-autoencoder pretraining to cine CMR with an unusually broad, openly reproducible benchmark. It compares against nnUNet and CNN baselines, tests external cohorts and measures scan–rescan consistency, which several earlier cardiac foundation models omitted. Its code and every pretrained and fine-tuned checkpoint are public under the MIT license. That openness has made it a reference point: CMRVision benchmarks directly against it. Its limits are also clear: pretraining is confined to UK Biobank cine imaging, and the systemic-disease and survival results await external validation.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.