PSIRNet Delivers Single-Acquisition Rapid LGE Reconstruction

PSIRNet is a physics-guided, end-to-end deep learning reconstruction that produces phase-sensitive inversion recovery (PSIR) late gadolinium enhancement (LGE) cardiac MRI from a single interleaved IR/PD acquisition over two heartbeats. The authors trained the model on a multi-institutional dataset of 800,653 slices from 55,917 patients acquired on 1.5T and 3T scanners, and the final model contains 845 million trainable parameters. Independent reader scoring found single-acquisition PSIRNet superior for dark-blood LGE and equivalent or superior for other LGE variants, while inference time dropped to roughly 100 msec per slice from more than 5 sec for the motion-corrected (MOCO) multi-average baseline.

PSIRNet is implemented as a variational, unrolled physics-guided network trained end-to-end to reconstruct PSIR images with surface coil correction directly from raw k-space input. Key technical points and evaluation details:

• •Training corpus split by patient produced 640,000 slices (42,822 patients) for training, with separate institutional validation and test sets to avoid data leakage.
• •Architecture: an unrolled variational network instantiation, total 845 million parameters, trained to map four 4D input tensors to a PSIR output tensor.
• •Metrics: quantitative comparisons used SSIM, PSNR, and NRMSE against MOCO PSIR references; qualitative scoring used a 5-point Likert scale with paired Wilcoxon tests and an equivalence margin of 0.25 Likert points.
• •Performance: clinical readers preferred or found equivalence; inference time is approximately 100 msec/slice. Hugging Face model card lists MIT license, research-use-only, and recommends GPUs with >=16GB memory, e.g., NVIDIA A100.

Free-breathing LGE acquisition currently relies on MOCO averaging of 8 to 24 repetitions to suppress motion and boost SNR, which increases scan time and reconstruction cost. Delivering diagnostic PSIR images from a single two-heartbeat acquisition addresses two bottlenecks at once: reduced patient time in the scanner and far faster reconstruction workflows. The scale of training data and cross-site evaluation strengthens generalizability compared with smaller academic studies, and Microsoft publishing a Hugging Face model card plus MIT license lowers friction for reproducibility and research adoption. This sits at the intersection of inverse-problem learning and clinically directed imaging AI and is likely to accelerate adoption of learned reconstruction in cardiology research.

Regulatory validation, prospective clinical trials, and robust external testing across vendor scanners and pathological subgroups will determine clinical translation. Implementation details such as handling off-nominal acquisition parameters and failure-mode detection are the immediate engineering gaps.