Clinical evaluation of deep learning-based acquisition and reconstruction techniques for accelerated 2D shoulder magnetic resonance imaging acceleration at 3 T.

PURPOSE: GE has recently developed Sonic DL MRI acceleration technology. We wished to (1) quantify scanner time reduction compared to the AIR Recon DL already in clinical use at our institution and (2) preliminarily evaluate any change in diagnostic findings between these methods. METHODS: A total of 30 outpatients undergoing routine shoulder MRI for shoulder pain were recruited. Each underwent routine AIR Recon DL and Sonic DL exams containing the same sequences. Three subspecialized musculoskeletal radiologists assessed image quality, graded pathology, and reported diagnostic confidence in a blinded fashion. Differences in image quality and confidence were compared using Wilcoxon signed-rank tests. RESULTS: Sonic DL reduced nominal acquisition time 38.7% compared to AIR Recon DL (6 min 15 s versus 10 min 12 s). Artifact was considered slightly worse on the routine examination (3.20 versus 3.37, P = 0.039); there was no significant difference in image quality, sharpness, or signal-to-noise ratio between routine and Sonic DL accelerated exams (P 0.171-0.861). Inter-reader agreement was broadly similar between routine-routine and routine-Sonic DL comparisons (median agreement 0.867 and 0.839, respectively). Within-reader analysis identified subacromial/subdeltoid bursitis and supraspinatus tendon tears more frequently on the routine protocol than on Sonic DL (P 0.049 and P 0.006, respectively). CONCLUSION: Sonic DL enabled a 38.7% reduction in nominal shoulder MRI acquisition time with broadly comparable image quality and inter-reader agreement in this pilot cohort. Diagnostic equivalence was not established; subacromial/subdeltoid bursitis and supraspinatus tears were potentially less conspicuous on Sonic DL. A larger powered validation study would be needed to fully evaluate diagnostic performance.