PSIRNet: Deep Learning-based Free-breathing Rapid Acquisition Late Enhancement Imaging

A multi-institutional research team has developed PSIRNet, a physics-guided deep learning model that dramatically accelerates cardiac MRI scans for late gadolinium enhancement (LGE) imaging. The 845-million-parameter network was trained on a massive dataset of 800,653 image slices from 55,917 patients collected across multiple sites between 2016 and 2024. Unlike traditional methods that require 8 to 24 motion-corrected signal averages taken over several minutes, PSIRNet produces diagnostic-quality images from a single acquisition spanning just two heartbeats, enabling patients to breathe freely during the procedure.

In clinical evaluations, two expert cardiologists performed independent qualitative assessments, scoring the AI-generated images against traditional motion-corrected PSIR references. For dark blood LGE variants, both readers rated PSIRNet reconstructions as superior, while for bright blood and wideband variants, one reader rated them superior and the other confirmed equivalence. The model also demonstrated remarkable speed improvements, reconstructing images in approximately 100 milliseconds per slice compared to more than 5 seconds for traditional methods.

The technology represents a significant advancement in medical imaging efficiency, potentially transforming cardiac care workflows. By reducing acquisition time 8- to 24-fold while maintaining diagnostic quality, PSIRNet could increase patient throughput, reduce costs, and improve accessibility to cardiac MRI diagnostics. The model's ability to work with free-breathing patients also enhances comfort and could expand the procedure to populations who struggle with breath-holding requirements.