PSIRNet: Deep Learning-based Free-breathing Rapid Acquisition Late Enhancement Imaging
A 845-million-parameter neural network cuts cardiac MRI scan time up to 24-fold by reconstructing diagnostic images from just two heartbeats of data.
The Thesis
Late gadolinium enhancement (LGE) MRI — the gold-standard scan for detecting scarred heart tissue — currently requires patients to hold their breath repeatedly while the scanner collects 8 to 24 averaged images to produce a clean result. PSIRNet replaces that lengthy process with a single two-heartbeat acquisition, using a physics-aware deep learning model trained on over 800,000 image slices from nearly 56,000 patients. The result: images that two independent cardiologists rated at least as good as — and often better than — the standard motion-corrected reference. The catch is that this is a retrospective study on existing data; prospective clinical trials and regulatory clearance remain ahead. If those hurdles clear, the technique could meaningfully shorten cardiac MRI exams, improve patient comfort, and increase scanner throughput.
Catalyst
Three things converged to make this feasible: massive multi-site clinical datasets (800,000+ labeled slices from real patients across eight years), sufficient GPU compute to train a near-billion-parameter model end-to-end, and mature physics-informed neural network architectures that embed MRI signal equations directly into the reconstruction pipeline. Two years ago, neither the labeled data nor the training infrastructure existed at this scale.
What's New
The prior standard is motion-corrected averaging (MOCO PSIR), which stacks 8–24 breath-hold or free-breathing acquisitions and uses image-registration algorithms to align them before averaging — a process that takes several minutes and is sensitive to arrhythmias and patient non-compliance. Earlier deep learning reconstruction efforts for cardiac MRI either worked on simpler image types, used much smaller datasets, or required multiple acquisitions as input. PSIRNet ingests a single interleaved inversion-recovery / proton-density acquisition lasting two heartbeats and outputs a fully corrected phase-sensitive image, cutting acquisition time by up to 24-fold while matching or exceeding MOCO image quality by cardiologist rating.
The Counter
This is a retrospective study — every image PSIRNet was tested on came from the same scanner ecosystem and patient population as its training data, even if institutions were held out. Real-world deployment will encounter scanner models, field strengths, coil configurations, and patient pathologies not in the training set, and reconstruction artifacts in those edge cases could be clinically dangerous in ways the Likert-scale study wouldn't catch. The 5-point cardiologist rating is a coarse instrument: two readers agreeing that images look 'equivalent' does not prove equivalent diagnostic accuracy for detecting small subendocardial infarcts or subtle fibrosis patterns. There is no downstream diagnostic outcome data — no comparison of missed infarcts, false-positive scars, or changed clinical decisions between PSIRNet and MOCO. Regulatory agencies like the FDA will almost certainly require prospective trials and outcome-linked validation before clearance, adding years and cost. Finally, the 845-million-parameter model requires meaningful GPU infrastructure that many community hospitals and lower-income health systems simply do not have at the point of scan.
Longs
- GEHC (GE HealthCare) — direct MRI scanner and AI reconstruction software revenue
- SIE (Siemens Healthineers ADR) — competing MRI platform; similar reconstruction AI integration potential
- PHG (Philips) — MRI systems and cardiac imaging software portfolio
- NVDA — GPU compute underpinning training and hospital inference servers
- ISRG — indirect; faster cardiac diagnostics support structural heart procedure pipeline
Shorts
- Motion-correction software vendors (e.g., companies whose sole product is MOCO post-processing for cardiac MRI — their workflow becomes unnecessary)
- MRI scanner manufacturers slow to integrate AI reconstruction (risk losing premium positioning to vendors who bundle PSIRNet-style models)
- Cardiac MRI service labs billing per-acquisition-average (revenue per exam could compress as scan time drops)
Enablers (Picks & Shovels)
- Siemens Healthineers and GE HealthCare MRI raw-data export pipelines (required for multi-site dataset curation)
- NVIDIA A100/H100 GPU clusters (845M-parameter model training at scale)
- PyTorch and JAX deep learning frameworks for physics-guided network design
- ISMRM open MRI reconstruction standards enabling reproducibility
- Multi-site clinical data-sharing agreements (the 55,917-patient dataset is itself an enabler others would need to replicate)
Private Watchlist
- Arterys (cardiac AI imaging, acquired by Tempus but technology lineage relevant)
- Cardio AI (cardiac MRI analytics startup)
- HeartVista (real-time MRI AI, cardiac focus)
- Hyperfine (portable MRI with AI reconstruction, public: HYPR)
Resources
The Paper
Purpose: To develop and evaluate a deep learning (DL) method for free-breathing phase-sensitive inversion recovery (PSIR) late gadolinium enhancement (LGE) cardiac MRI that produces diagnostic-quality images from a single acquisition over two heartbeats, eliminating the need for 8 to 24 motion-corrected (MOCO) signal averages. Materials and Methods: Raw data comprising 800,653 slices from 55,917 patients, acquired on 1.5T and 3T scanners across multiple sites from 2016 to 2024, were used in this retrospective study. Data were split by patient: 640,000 slices (42,822 patients) for training and the remainder for validation and testing, without overlap. The training and testing data were from different institutions. PSIRNet, a physics-guided DL network with 845 million parameters, was trained end-to-end to reconstruct PSIR images with surface coil correction from a single interleaved IR/PD acquisition over two heartbeats. Reconstruction quality was evaluated using SSIM, PSNR, and NRMSE against MOCO PSIR references. Two expert cardiologists performed an independent qualitative assessment, scoring image quality on a 5-point Likert scale across bright blood, dark blood, and wideband LGE variants. Paired superiority and equivalence (margin = 0.25 Likert points) were tested using exact Wilcoxon signed-rank tests at a significance level of 0.05 using R version 4.5.2. Results: Both readers rated single-average PSIRNet reconstructions superior to MOCO PSIR for dark blood LGE (conservative P = .002); for bright blood and wideband, one reader rated it superior and the other confirmed equivalence (all P < .001). Inference required approximately 100 msec per slice versus more than 5 sec for MOCO PSIR. Conclusion: PSIRNet produces diagnostic-quality free-breathing PSIR LGE images from a single acquisition, enabling 8- to 24-fold reduction in acquisition time.