# Clinical Application of Deep Learning for Spine MRI Interpretation: A Multicenter Evaluation of Artificial-Intelligence-Assisted versus Manual Reading on Diagnostic Agreement with the Reference Standard

## TL;DR

This study introduces an AI system for real-time MRI spine analysis, showing high accuracy and consistency compared to manual methods.

## Contribution

Lumbar VNet Pro (LVP) is the first real-time AI system integrated into MRI hardware for lumbar spine diagnostics.

## Key findings

- LVP achieved 100% recognition accuracy in internal testing with 97% consistency against manual assessments.
- AI-assisted methods outperformed manual approaches in diagnosing spinal pathologies with AUC > 0.95.
- LVP improved positioning accuracy and reduced interobserver variability in MRI spine diagnostics.

## Abstract

Lumbar spine diseases substantially impact the patients’ quality of life, necessitating accurate and efficient diagnostic tools. This study presents Lumbar VNet Pro (LVP), the first real-time artificial-intelligence (AI)-assisted system embedded within MRI hardware for lumbar spine analysis, integrating deep learning with MRI. LVP was trained on 2,453 MRI datasets and validated both internally and externally across multiple centers. During the training (1,848 MRI datasets) and validation (605 MRI datasets), LVP exhibited outstanding performance in localization (Dice = 0.93), segmentation (Dice = 0.92), labeling (identification rate = 0.90), and timeliness (average inference time = 1.1 s). Following the successful construction of LVP, we conducted comprehensive testing through both internal and external multicenter evaluations. Internal testing involving 100 patients indicated that the recognition accuracy of LVP was as high as 100%, and the consistency between the LVP assessment and the manual assessment using the gold standard reached 97%. In external testing involving 1,522 patients, LVP’s diagnostic performance was compared to those of manual and human–machine-assisted methods. The AI-assisted approaches demonstrated better performance across multiple spinal pathologies, including lumbar disc herniation, spinal canal stenosis, and lateral recess stenosis, with area under the receiver operating characteristic curve values >0.95 for deep learning/human–machine approaches and >0.90 for the fully manual approach. The real-time integration of LVP with MRI scanning improved positioning accuracy and reduced interobserver variability, supporting its potential as an adjunct tool for enhancing MRI-based spine diagnostics. However, further studies are warranted to assess its generalizability across diverse clinical settings.