Source-linked AI summary

MR-JEPA: A General Purpose Video Foundation Model for Cardiac MRI

Athira J. Jacob, Puneet Sharma, Dorin Comaniciu, Daniel Rueckert

arXiv:2608.30975v1cs.CVcs.AI

TL;DR

Most CMR deep learning methods discard sequential context by processing individual 2D slices, motivating a unified model for temporal cine and spatial multi-sequence data. MR-JEPA extends LeJEPA to 3D spatiotemporal CMR with tubelet embeddings, masking, and 2D CMR initialization, then evaluates a frozen encoder across six tasks. It outperforms compared methods on all five regression tasks and remains competitive for disease classification, while its evaluation is limited by single-vendor pretraining and restricted external validation.

  • Problem

    Most CMR deep learning methods process individual 2D slices, disregarding sequential context, while adapting video models is difficult because of limited medical video, heterogeneous sequences, and multi-view integration needs.

  • Method

    MR-JEPA extends LeJEPA to 3D spatiotemporal CMR using tubelet embeddings, spatiotemporal masking, and initialization from a 2D CMR foundation model, with gated attention over frozen-encoder views.

  • Results

    MR-JEPA achieves the best performance on all five regression tasks and remains competitive for disease classification, including LV EF MAE 4.79% and GLS MAE 1.87.

  • Takeaways & Limitations

    The results support a unified encoder for multi-view utilization of diverse CMR sequences in cardiac quantification and diagnosis.

  • Takeaways & Limitations

    Pretraining uses single-vendor data; most downstream tasks come from one pretraining center, and RV EF and strain references derive from a validated deep learning model rather than manual annotations.

Abstract

from arXiv · show

Cardiac magnetic resonance imaging (CMR) produces rich sequential data such as temporal cine videos and spatial LGE/mapping stacks, yet most deep learning approaches process individual 2D slices, discarding this context. We present MR-JEPA, a self-supervised video foundation model for CMR that extends LeJEPA to 3D spatiotemporal inputs through tubelet tokenization, spatiotemporal masking augmentation, and initialization from a 2D CMR foundation model. Unlike prior CMR video models limited to cine data, MR-JEPA is pretrained on multi-sequence data (cine, LGE, mapping) from 10,505 patients across two centers without annotations. We evaluate the frozen encoder on six downstream tasks using a unified multi-view gated attention architecture: LV ejection fraction, RV ejection fraction, three myocardial strains (GLS, GCS, GRS), and four-class disease detection. MR-JEPA outperforms other compared methods on all five regression tasks, including both a domain-specific CMR model pretrained on more data with text supervision and a natural-video foundation model, achieving an LV EF MAE of 4.79% (r =0.764) and a GLS MAE of 1.87 (r=0.805), with 21-27% MAE reductions over baselines on strain tasks. For disease detection, MR-JEPA achieved a macro AUG of 0.868, remaining competitive with the domain-specific baseline despite using a fully self-supervised pretraining objective. These results demonstrate the potential of a unified video encoder for robust, multi-view utilization of diverse CMR sequences in clinical cardiac quantification and diagnosis.

1 Introduction

CMR contains rich temporal, spatial, and tissue-characterization information, but most deep learning methods analyze individual 2D slices. MR-JEPA addresses this gap with a self-supervised video foundation model designed for heterogeneous multi-sequence CMR.

  • CMR combines cine, LGE, and parametric mapping sequences containing temporal and spatial information relevant to cardiac function, mechanics, and disease characterization.
  • Most CMR deep learning methods process individual 2D slices, disregarding sequential context in cine and multi-slice acquisitions.
  • MR-JEPA is a fully self-supervised CMR video foundation model pretrained on multi-sequence data without annotations.
  • MR-JEPA extends LeJEPA with tubelet embeddings, spatiotemporal masking, and initialization from a pretrained 2D CMR foundation model.
  • MR-JEPA outperforms natural-video and prior CMR-specific foundation models on all regression tasks while remaining competitive for disease classification.

2 Related Work

Self-supervised visual representation learning has expanded from images to video, while cardiac MRI foundation models have explored 2D, masked-autoencoder, and multimodal approaches. Existing CMR video methods remain limited by cine-focused inputs or dependence on paired clinical reports.

  • Image-based self-supervised methods learn invariant representations from augmented views using contrastive or self-distillation objectives.
  • Video self-supervised learning includes masked autoencoders that reconstruct masked patches and JEPA methods that predict latent representations.
  • Prior cardiac MRI foundation models include a 2D DINO-based model and CineMA, which uses masked-autoencoder pretraining on cine images.
  • Shad et al. proposed multimodal foundation-model pretraining on cine MRI and clinical reports, but such approaches require paired report-image data.

3 Method

MR-JEPA extends LeJEPA into a unified spatiotemporal encoder for diverse CMR sequences. A frozen encoder produces view embeddings that gated attention aggregates for cardiac function regression and disease classification.

  • LeJEPA trains a single shared encoder to learn invariant representations across multiple augmented views without teacher-student components.
  • MR-JEPA adapts LeJEPA to CMR video using 3D tubelet embeddings, spatiotemporal masking, and initialization from a 2D CMR foundation model.
  • The Vision Transformer jointly encodes spatial and temporal tubelets, while its [CLS] output is projected for the pretraining loss.
  • The frozen encoder is evaluated on five cardiac-function regression tasks covering LV EF, RV EF, GLS, GCS, and GRS.
  • A sixth task classifies patients as normal, DCM, HCM, or IHD using cine views and an LGE view.
  • Gated attention aggregates projected [CLS] embeddings from multiple views before a task-specific linear head produces regression or classification outputs.

4 Data and Experiments

MR-JEPA is pretrained and evaluated within a frozen-encoder, multi-view framework using diverse CMR data and matched baselines. Experiments cover cardiac function, disease classification, and analyses of key pretraining choices.

  • Pretraining uses 160,172 CMR clips from 10,505 patients across two clinical centers, acquired on 1.5T and 3T Siemens scanners.
  • RV EF and strain tasks use Center 1 data with reference measurements generated by a previously validated CMR analysis pipeline.
  • Disease labels are normal, IHD, HCM, or DCM, with inputs comprising three cine videos and one whole-heart LGE stack.
  • Auxiliary analyses assess cine-view classification and LGE scar detection to evaluate representation quality across temporal and spatial sequences.
  • Pretraining uses a 27.0M-parameter ViT-S/16 for 300 epochs with eight input frames, multi-scale crops, and two-stage resolution training.
  • All methods use frozen encoders with the same gated-attention downstream architecture and identical downstream data and splits for fair comparison.
  • MR-JEPA is compared with a text-supervised CMR model pretrained on 293K OMR videos and V-JEPA2, a natural-video JEPA model.

5 Results

MR-JEPA achieved the strongest performance across the five cardiac function regression tasks, while remaining competitive for disease classification. Ablations identified initialization, spatiotemporal masking, and shorter clips as important design choices.

  • Cardiac Function Regression: MR-JEPA achieved the best MAE and Pearson correlation on all five cardiac function regression tasks.
  • LV Ejection Fraction: 4.79% LV EF MAE and r=0.764 outperformed Shad et al. and V-JEPA2 on the Kaggle dataset.MR-JEPA also showed minimal systematic bias of 0.22%.
  • RV Ejection Fraction: 8.67% RV EF MAE and r=0.506 were the best reported results, improving over Shad et al. by 0.68 percentage points.All methods had wider limits of agreement than for LV EF.
  • Myocardial Strain: Strain MAEs were 1.87 for GLS, 2.39 for GCS, and 5.33 for GRS, representing 15-27% reductions versus the strongest baseline.Correlations ranged from 0.726 to 0.805, with absolute systematic bias below 0.97.
  • Cardiac Disease Detection: MR-JEPA reached macro AUC 0.868 for disease classification, below Shad et al.'s 0.882 but above V-JEPA2's 0.741.It achieved the highest normal-class one-vs-rest AUC at 0.889 and remained within 0.01-0.03 AUC of Shad et al. across disease classes.
  • Model Analysis: Removing 2D-CMR initialization or spatiotemporal masking reduced scar-detection accuracy from 0.850 to chance level, while 16-frame clips lowered view-classification accuracy from 0.932 to 0.841.The 8-frame configuration also required fewer optimization epochs, 103 versus 190.

6 Conclusions

MR-JEPA extends LeJEPA to 3D spatiotemporal cardiac MRI and learns representations from multi-sequence data without annotations. It outperformed compared methods on five regression tasks and remained competitive for disease detection, while the study retains important data and evaluation scope limitations.

  • Conclusions: MR-JEPA extends LeJEPA to 3D spatiotemporal inputs and uses spatiotemporal masking with 2D-CMR foundation-model initialization.It was pretrained on 160K multi-sequence clips from 10,505 patients without annotations.
  • Conclusions: MR-JEPA achieved state-of-the-art performance on LV EF, RV EF, GLS, GCS, and GRS regression tasks against domain-specific and natural-video baselines.
  • Conclusions: MR-JEPA remained competitive on disease detection despite less pretraining data and a simpler unsupervised objective.
  • Limitations: The study is limited by single-vendor pretraining, center-limited downstream evaluation for most tasks, model-derived RV EF and strain ground truth, and cine-dominated data.The authors identify more mapping data, mapping-specific tasks, and comparisons with task-specific state-of-the-art models as future work.
Loading 2608.30975v1…