Examining the Effect of Deep Learning-Based Image Reconstruction on Accelerating Shoulder Magnetic Resonance Imaging (MRI) and Its Impact on Image Quality.

Background Prolonged scan time remains the main obstacle to increasing magnetic resonance imaging (MRI) throughput. The advent of artificial intelligence brings forth opportunities to accelerate MRI examinations. Purpose This study compares the image quality of standard MRI versus accelerated MRI with deep learning-based image reconstruction (DLR) for shoulder MRI studies. Materials and methods Forty-nine subjects were prospectively enrolled and underwent both standard and accelerated axial proton density fat-saturated (PD FS) shoulder MRIs using a 1.5T scanner (Philips Ingenia 1.5T). Two blinded musculoskeletal radiologists independently evaluated paired datasets to assess the anatomic conspicuity of specific structures (labrum, rotator cuff footprint, cartilage, long head of the biceps tendon/rotator interval), artifacts, and overall image quality. A 5-point scale was employed, where 1 indicated the standard MRI was markedly superior and 5 indicated the accelerated MRI was markedly superior. The reduction in scan time was recorded; inter-reader variability was also analyzed. Results The DLR protocol reduced scan duration by 20.2% on average, shortening acquisition time from 184 seconds to 148 seconds. Mean scores for anatomic conspicuity ranged from 3.0 to 3.2, and mean scores for artifacts and overall image quality were 3.0 and 3.2, respectively. The Wilcoxon signed-rank test revealed statistically significant differences (p<0.001) for most categories, except for "Artifacts" as assessed by one reader. Inter-reader agreement was poor, with Cohen's kappa ranging from 0.086 to 0.183 and prevalence-adjusted bias-adjusted kappa (PABAK) scores ranging from 0.063 to 0.404. Conclusion DLR-based acceleration significantly reduces scan time while maintaining diagnostic image quality, presenting a clinically feasible and efficient solution for routine shoulder MRI.