ORION-CMR: On-scanner Reporting with Integrated Foundation Model for End-to-End Cardiac MRI Analysis and Interpretation
Omer Burak Demirel, Kelly K. Horst, Alessio Perazzolo, Elisa Bruno, Kenan Kaya, Rongzhen Ouyang, Enas Ahmed, Jouke Smink, et al.
eess.IV · cs.CV · cs.LG · physics.med-ph
Quick verdict
Scanner-native CMR foundation model with real clinical validation incl. reports; strong on EMIDEC, small single-center hold-out (n=68) tempers clinical claims.
Summary
The paper presents ORION-CMR, a self-supervised (DINO-style, ViT-B) CMR foundation model pretrained on ~12.9 million images from 9,258 multi-vendor studies at Mayo Clinic, designed to run entirely on an MRI scanner (Philips 1.5T Ambition X with RTX A6000 GPU). A single frozen encoder supports sequence classification, cine short-axis and LGE segmentation, LGE/binary/multiclass disease classification, and LLM-based report generation via a locally deployed Qwen2.5-14B, producing a report in ~90 seconds per study. On public benchmarks it sets new EMIDEC results (LGE classification 0.930 vs 0.920 SoTA; myocardium Dice 0.915, scar Dice 0.799) and beats CMR-FM on ACDC classification (0.820 vs 0.700), though it stays below published ACDC SoTA for both classification (0.820 vs 0.960) and myocardium segmentation (0.878 vs 0.922). In a 68-subject single-center hold-out (25 normal, 7 CHD, 15 DCM, 21 MI), AUC reached 0.960 for normal-vs-abnormal and 0.845-0.960 per class; expert readers judged 81.4% of the 68 generated reports concordant, with moderate inter-reader agreement (kappa 0.41-0.61). Main limitations: small single-center validation with only 7 CHD cases, ventricular metrics validated against clinical reports rather than expert contours, unnamed statistical tests, and no code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● 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
- ORION-CMR is the first clinically evaluated, scanner-native, end-to-end CMR foundation model, pretrained on 12,896,733 images from 9,258 multi-vendor studies, completing analysis plus report generation in about 90 seconds per study.
- A single self-supervised ViT-B8 encoder outperforms supervised baselines and CMR-FM across public benchmarks, setting new SoTA on EMIDEC for LGE classification (0.930), myocardium (Dice 0.915) and scar (0.799) segmentation.
- In a 68-subject single-center hold-out, the system achieves AUC 0.960 (CI 0.906-0.998) for normal-vs-abnormal classification and 0.845-0.960 per-class AUC for multiclass disease classification.
- Reports generated by a locally deployed Qwen2.5-14B from a deterministic structured representation agree with expert interpretation in 81.4% of 68 cases, with only 4.9% judged discordant.
- The complete privacy-preserving pipeline runs on-scanner (Philips 1.5T Ambition X, RTX A6000) without external network communication, exporting the report as a DICOM series.
Strengths
Genuine end-to-end scanner-native deployment with clinical evaluation of reports
The full pipeline runs on a Philips 1.5T Ambition X with RTX A6000 (§2.6): sequence classification triggers downstream tasks as series arrive, and the local Qwen2.5-14B report is exported as DICOM with no external network. Critically, the clinical evaluation covers the generated reports themselves (§3.2), not just image metrics - rare for foundation-model papers.
Architecture constrains LLM hallucination by design
Reports are synthesized only from a deterministic structured representation (ventricular metrics interpreted via HHC 2024 reference ranges plus classifier outputs, §2.4.4), separating image interpretation from language generation. Result: 81.4% concordant and only 4.9% discordant ratings across 68 reports (§3.2).
Single frozen encoder with honest transfer protocol; real SoTA on EMIDEC
Sequence classification is a practical standout
Transparent operational profiling and candid limitations
Weaknesses
Below published SoTA on ACDC, underemphasized in the abstract
Quantitative core of every clinical report is unvalidated against contours
Very small, imbalanced single-center hold-out
Under-specified statistics and possibly under-tuned baselines
No code, weights, or data release; generalization gap undiscussed
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
ORION-CMR (arXiv 2609.23950) claims to be the first clinically evaluated, scanner-native, end-to-end CMR foundation model. A single DINO-style self-supervised ViT-Base encoder - pretrained on 12,896,733 slices from 9,258 multi-vendor (GE/Siemens) studies collected at Mayo Clinic 2017-2025 - supports five task families: 11-way sequence classification, cine short-axis and LGE segmentation, LGE plus binary/multiclass disease classification, and LLM-drafted reporting via a locally deployed Qwen2.5-14B-Instruct. Headline claims: new state of the art on EMIDEC for LGE classification and scar segmentation; superiority over supervised baselines and the CMR-FM foundation model; clinical AUC 0.960 (normal vs abnormal) in a 68-subject hold-out; and 81.4% expert-rated concordance of generated reports, all in ~90 seconds per study on a Philips 1.5T Ambition X with an RTX A6000, without external network calls.
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 ORION-CMR?
ORION-CMR is a cardiac MRI (CMR) foundation model developed by Mayo Clinic and Philips researchers. It runs entirely on the MRI scanner: it classifies image sequences, segments heart chambers, detects scar (late gadolinium enhancement), classifies disease, and drafts a clinical report using a locally hosted language model (Qwen2.5-14B), all in about 90 seconds without sending data off the scanner.
How accurate is ORION-CMR?
On public benchmarks it set new best results on EMIDEC (0.930 for LGE classification; Dice 0.915 for myocardium and 0.799 for scar) and beat the prior CMR-FM model on ACDC (0.820 vs 0.700), though it remained below the best published ACDC results (0.960). In a 68-patient clinical test, AUC was 0.96 for normal-vs-abnormal, and expert radiologists judged 81.4% of its generated reports concordant with their own interpretation.
Has ORION-CMR been validated in real patients?
Partially. Validation was retrospective, on 68 held-out patients from a single center (Mayo Clinic) covering normal exams, congenital heart disease, dilated cardiomyopathy and myocardial infarction; three radiologists rated 81.4% of generated reports as unlikely to change treatment decisions. Larger multi-center, prospective validation has not yet been performed, and ventricular measurements were compared against clinical reports rather than expert-drawn contours.
Does ORION-CMR use ChatGPT or cloud AI for its reports?
No. It uses Qwen2.5-14B-Instruct running locally through Ollama on an NVIDIA RTX A6000 GPU at the scanner, and the finished report is exported as a DICOM series with no external network communication, which preserves patient privacy.
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.
- Figures 1-3 (pipeline schematic, report examples with reader scores, concordant/discordant cases) are referenced but not analyzable from extracted text; actual report quality cannot be independently assessed.
- The abstract describes 'a multi-vendor cohort of 68 subjects' while section 2.2 describes a 426-subject cohort with a 68-subject hold-out; this review follows the methods section.
- Potentially supplementary details (hyperparameter search ranges, LLM prompts, QC criteria) are not in the extracted text and may exist in an appendix not captured.
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.