Prospective evaluation of artificial intelligence (AI) in lumbar spine magnetic resonance imaging (MRI) workflow: from deep learning (DL)-enhanced accelerated acquisition to simultaneous vision-language model (VLM)-based automated report generation

Abstract

Objectives: To evaluate the diagnostic interchangeability of DL-enhanced accelerated lumbar (L)-spine magnetic resonance imaging (MRI) with conventional imaging and to assess the diagnostic agreement and feasibility of vision-language-model (VLM)-based automated reporting. Methods: The Institutional Review Boards oftwo participating institutions approved this prospective study. Seventy patients were enrolled from these two institutions. All the participants underwent both conventional and accelerated L-spine MRI during the same session, resulting in 140 MRI scans. Quantitative analyses included signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR), whereas qualitative image quality assessments were conducted by four radiologists blinded to the scan type and patient information. The interchangeability between conventional and accelerated MRI with DL-based enhancement protocols was evaluated for five key pathologic findings. Automated structured reports were generated using a commercially available VLM-based spine interpretation software and compared with radiologist consensus reports. Statistical analyses were performed, with p < 0.05 considered statistically significant. Results: Accelerated L-spine MRI with DL-based enhancement reduced the acquisition time by approximately 80-86% when compared with conventional MRI, while maintaining diagnostic interchangeability. Quantitative analyses revealed superior SNRs and CNRs, and qualitative evaluations supported comparable image quality. Automated reporting demonstrated substantial to almost perfect agreement across key pathologies. Conclusions: DL-enhanced accelerated MRI produced high-quality diagnostic images within 2 min, and VLM-based automated reporting demonstrated strong agreement with the radiologists. These findings provide prospective evidence supporting the clinical feasibility of integrating AI into both the acquisition and interpretation workflows in L-spine MRI, with the potential to enhance the efficiency, consistency, and scalability of musculoskeletal imaging.