Deep learning-based super-resolution reconstruction for improved image quality in shoulder MRI

Purpose To compare the image quality of shoulder Magnetic Resonance Imaging (MRI) using three imaging acquisition or reconstruction approaches—Parallel Imaging (PI), Compressed Sensing (CS), and Deep Learning-Based Super-Resolution Reconstruction (SR)—and to evaluate reader-assessed image quality and preference. Methods A retrospective study included 60 patients (27 males, 33 females; mean age 53.67 ± 12.11 years). Shoulder MRI was performed with PI, CS and SR (reconstructed from CS data) across six sequences. Quantitative and qualitative assessments were conducted by experienced observers. Three radiologists selected clinically preferred sequences after a 2-week washout. Consistency and inter-group comparisons were statistically analyzed. Results The quantitative and qualitative assessment between observers showed good consistency. The SR group showed significantly higher quantitative measurements across all six sequences of structures [e.g. SNR: Supraspinatus tendon SR 14.03 (10.60, 20.26), CS 11.87 (8.39, 15.96), PI 8.75 (5.18, 13.51), P < 0.001] than PI and CS. SR images generally showed lower edge rise distance (ERD) values, significant overall differences among PI, CS, and SR ( P < 0.01), Qualitative assessment revealed that SR overall differences among PI, CS and SR, with statistically significant differences ( P < 0.001). Furthermore, three reader preference results indicated a preference for SR images that all readers preferred SR images (86 %, 92 %, and 93 %). Conclusion Compared to PI and CS, SR significantly improves image quality in shoulder MRI without increasing acquisition time. SR enhances the visualization of fine anatomical structures, providing improved image quality and supporting its potential application in clinical shoulder MRI.