Lumbar spine MRI: comparison of novel deep learning algorithm and conventional sequences on 1.5T

Purpose or Learning Objective To prospectively compare quantitative and subjective image quality and scanning time between a new deep learning-based reconstruction (DLR) algorithm and standard MRI protocol of the lumbar spine. Methods or Background Thirty-one healthy volunteers underwent 1.5T MRI examination of lumbar spine from April to September 2021. Protocol acquisition comprised sagittal T1- and T2-weighted and short-tau inversion recovery (STIR) images and axial T2-weighted images. All sequences were acquired with both DLR and standard protocols. The quantitative image analysis with signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) was performed by a radiologist drawing regions of interest (ROIs) on the fourth lumbar vertebral body and in the intervertebral disc (L4/5); the qualitative image analysis between the two protocols was performed by two radiologists in blind. Scanning times were also recorded and compared. Results or Findings The SNR of DLR algorithm was higher in all sequences in the vertebrae and discs compared to standard images (all p<0.001). The CNR of the DLR algorithm was superior to conventional T2-weighted images (p<0.001), whereas no significant differences were reported for T1-weighted (p=0.67) and STIR images (p=0.40). Qualitative analysis showed that DLR had greater overall quality in all sequences (all p<0.001), with an inter-rater agreement of 0.69 (0.63-0.76). Total protocol scanning time was lower in DLR compared to standard protocol (average acquisition time 6.33 vs 13.06 minutes, p<0.001), resulting in acquisition time reduction of 49%. Conclusion DLR algorithm applied to 1.5T MRI is a feasible method for lumbar spine imaging providing morphologic sequences with a higher SNR, CNR and image quality, compared with standard protocol with a significant scanning time reduction. Limitations Small population sample and only one district analysed.