Clinical evaluation of deep learning accelerated lumbar T2-weighted and fat-suppressed MRI sequences.

Abstract

OBJECTIVE: To prospectively compare the scanning efficiency, image quality, and diagnostic performance of deep learning (DL)-based turbo spin-echo sequences (TSE-DL, TSE-DL-FS) with conventional sequences (TSE-SD, TSE-FS) and dual-echo Dixon water–fat separation sequences (TSE-Dixon) in lumbar spine MRI, providing evidence for clinical protocol optimization. METHODS: A total of 71 patients with lumbar spine disorders were prospectively enrolled. Each underwent five MRI sequences. The DL sequences were reconstructed via a U-Net–based network. Image evaluation was performed by two blinded radiologists using 4-point and 5-point Likert scales. Intergroup comparisons were conducted with the Kruskal–Wallis H test and Bonferroni post-hoc corrections. Signal-to-noise ratio (SNR) was measured using regions of interest (ROI) and compared via one-way ANOVA and Bonferroni tests. Disc–cerebrospinal fluid ratio (DCFR) was calculated and correlated with Pfirrmann grades via Spearman analysis. RESULTS: The total scan time of TSE-DL and TSE-DL-FS (118 s) was significantly shorter than that of conventional TSE-SD + TSE-FS (202 s) and TSE-Dixon (145 s), improving efficiency by 41.6% and 18.6%, respectively. Qualitative and quantitative image scores of TSE-DL + TSE-DL-FS were comparable to conventional sequences (P > 0.05), but superior to TSE-Dixon (P < 0.05). Although TSE-Dixon achieved the most homogeneous fat suppression, its SNR was the lowest. DCFR showed a negative correlation with Pfirrmann grade in all sequences, strongest in TSE-DL (r = − 0.655, P < 0.001). CONCLUSION: The combination of DL-reconstructed TSE-DL and TSE-DL-FS sequences can markedly reduce lumbar MRI acquisition time while maintaining diagnostic image quality. Moreover, TSE-DL demonstrates potential benefits in the quantitative evaluation of disc degeneration. This highlights its promise for broader clinical adoption.