MR-JEPA: A General Purpose Video Foundation Model for Cardiac MRI
Athira J. Jacob, Puneet Sharma, Dorin Comaniciu, Daniel Rueckert
cs.CV · cs.AI
Quick verdict
Self-supervised CMR video FM tops both baselines on all 5 regression tasks; convincing but untested significance, and center-overlap leakage unresolved.
Summary
MR-JEPA is a self-supervised video foundation model for cardiac MRI that extends LeJEPA to 3D spatiotemporal inputs via tubelet tokenization, spatiotemporal masking, and initialization from a 2D CMR foundation model. It was pretrained without annotations on 160,172 multi-sequence clips (77% cine, 23% LGE, 0.4% mapping) from 10,505 patients at two Siemens-only centers. With the encoder frozen and a gated-attention multi-view head, it achieves the best MAE and Pearson r among the two compared baselines on all five regression tasks: LV EF MAE 4.79% (r=0.764) on the external Kaggle test set, RV EF 8.67 (r=0.506), GLS 1.87 (r=0.805), GCS 2.39, GRS 5.33 — 21-27% MAE reductions over the strongest baseline on strain tasks. On four-class disease detection it reaches macro AUC 0.868, slightly below the text-supervised baseline (0.882). Ablations show removing 2D initialization or masking collapses scar probing to chance. Key limitations: reference standards for four tasks come from an automated pipeline, downstream evaluation (except LV EF) uses data from a pretraining center, no significance tests or seed variance are reported, only two baselines are compared, and no code or data release is mentioned.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- LeJEPA's heuristics-free joint-embedding objective can be extended to 3D CMR video via tubelet tokenization, spatiotemporal masking, and initialization from a 2D DINOv3-based CMR foundation model (Section 3).
- With a frozen encoder and gated-attention aggregation, MR-JEPA achieves the best MAE and Pearson r among the two compared baselines on all five regression tasks, including LV EF MAE 4.79% (r=0.764) on an external test set and 21-27% MAE reductions on strain tasks (Table 1).
- On four-class disease detection MR-JEPA reaches macro AUC 0.868, competitive with but below the text-supervised Shad et al. baseline (0.882) despite no text supervision, 5/6 fewer... (as stated: fewer) pretraining clips and a smaller model (Table 2).
- Ablations show 2D CMR initialization and spatiotemporal masking are necessary for transfer, with scar probing falling to chance without either (Table 3).
- Pretraining on multiple CMR sequences (vs cine-only) may improve cross-contrast representations — asserted but not directly ablated.
Strengths
Multi-sequence pretraining scope
160,172 clips from 10,505 patients spanning cine, LGE and mapping (Section 4.1), unlike prior CMR video models (Shad et al., CineMA) restricted to cine; no text supervision required.
External validation for LV EF
Evaluated on the independent multi-center Kaggle Data Science Bowl test set (n=427) with a clear margin: MAE 4.79 / r=0.764 vs 6.07/0.614 (Shad et al.) and 6.70/0.189 (V-JEPA2), plus near-zero bias 0.22 (Table 1).
Honest classification reporting
Informative ablations (Table 3)
Uniform comparison protocol
Weaknesses
Automated reference standards for 4/5 regression tasks
No statistical inference or seed variance
Multi-sequence headline claim never ablated
Only two baselines, frozen-encoder only
Reproducibility signals are weak
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
MR-JEPA (Sections 1-3) makes three claims: (1) LeJEPA's heuristics-free joint-embedding objective can be ported from 2D natural images to 3D CMR video via tubelet tokenization, spatiotemporal masking of local crops, and initialization from a 2D DINOv3-based CMR foundation model; (2) a fully self-supervised video encoder pretrained on 160,172 clips from 10,505 patients (77% cine, 23% LGE, 0.4% mapping; two Siemens-only centers) yields the best MAE and Pearson r on all five regression tasks (LV/RV ejection fraction, GLS/GCS/GRS strain) against two baselines — the text-supervised Shad et al. model and natural-video V-JEPA2; (3) it remains competitive on four-class disease detection (macro AUC 0.868 vs 0.882). All claims are explicitly scoped to frozen-encoder evaluation under one shared gated-attention head, and the authors concede in Limitations that task-specific SOTA comparisons are absent.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is MR-JEPA?
MR-JEPA is a self-supervised video foundation model for cardiac MRI. It extends the LeJEPA joint-embedding objective to 3D spatiotemporal input using tubelet tokens, spatiotemporal masking, and initialization from a 2D cardiac MR foundation model, and was pretrained without any annotations on 160,172 CMR clips (cine, LGE, mapping) from 10,505 patients at two centers equipped with Siemens scanners.
How accurate is MR-JEPA at estimating ejection fraction and strain?
With its encoder frozen, it reports the lowest error among the two compared baselines on all five regression tasks: LV ejection fraction MAE 4.79% (r=0.764) on an external Kaggle test set, RV EF MAE 8.67 (r=0.506), GLS 1.87 (r=0.805), GCS 2.39, and GRS 5.33, i.e., 21-27% lower MAE than the strongest baseline on the strain tasks.
Does MR-JEPA beat previous cardiac MRI foundation models?
On the five function/strain regression tasks it outperforms both compared baselines, including a text-supervised CMR video model pretrained on more data. On four-class disease detection it is slightly behind that text-supervised baseline (macro AUC 0.868 vs 0.882) while clearly ahead of the natural-video baseline (0.741).
Is the MR-JEPA code or data publicly available?
The paper does not mention any code, weight, or data release. The pretraining and most evaluation data are proprietary clinical scans from two centers (Siemens scanners only); only the LV EF evaluation dataset (Kaggle Data Science Bowl) is public.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Loss equations (Eq. 1-2) and some inline math are OCR-garbled in the extraction; numeric values in Tables 1-3 remain legible and were used as reported.
- The Conclusions cite an 'embedding analysis' (temporal/spatial separation) that is absent from the provided text; it may reside in supplementary material not included in the extraction.
- Figure 1 is present only as a caption in the text extraction; its visual content could not be reviewed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.