AI-assisted MRI workflow enables real-time spine positioning and analysis in multicenter study

Background

Magnetic resonance imaging is a core tool for evaluating lumbar spine disorders, including disc herniation, spinal canal stenosis, and related degenerative conditions. In routine practice, however, scan planning and image interpretation still depend heavily on operator experience. Technologists typically determine scan planes manually, and small differences in positioning may influence downstream measurements and affect diagnostic consistency. Although artificial intelligence has shown promise in medical image analysis, most current systems are used for post-processing after image acquisition. As a result, real-time integration of AI into the MRI scanning process itself has remained limited. The present study was developed to address that gap by moving AI support upstream into the acquisition stage.

Research Progress

To support real-time MRI assistance, the research team developed Lumbar VNet Pro, a V-Net–based deep-learning framework that can automatically identify vertebral structures and intervertebral discs while generating scanning guidance during MRI acquisition(Figure 1). The model was trained on 2,453 MRI datasets and showed strong performance in core technical tasks, including anatomical localization with a Dice coefficient of 0.93, disc segmentation with a Dice coefficient of 0.92, and an average inference time of about 1.1 seconds, indicating that the system could operate within real-time clinical constraints.

The system was then evaluated in a multicenter study involving 1,522 patients. In this clinical testing phase, the investigators compared fully automated deep-learning analysis, human–machine collaborative assessment, and fully manual interpretation. The results showed that AI-assisted approaches improved MRI positioning accuracy and reduced interobserver variability across several structural measurements. In diagnostic tasks involving lumbar disc herniation, spinal canal stenosis, and lateral recess stenosis, the AI-assisted methods achieved area-under-the-curve values above 0.95, while the fully manual approach remained above 0.90. These findings suggest that embedding AI into the MRI acquisition workflow may improve both imaging consistency and downstream diagnostic performance(Video 1).

Future Prospects

The researchers note that the significance of this work lies not only in improving image analysis performance, but also in demonstrating a different model for clinical AI deployment. Rather than functioning solely as an offline reading tool, the LVP system allows AI to participate during image acquisition through a closed-loop process that connects scanning, analysis, and feedback. In principle, this type of device-embedded AI could support more standardized scan planning, reduce operator dependence, and improve workflow efficiency across institutions. The authors suggest that, with further validation and broader deployment, similar systems may contribute to future imaging platforms that incorporate automated protocol generation, real-time quality control, and more integrated decision support.

Video 1 demonstration of the system: https://mp.weixin.qq.com/s/iVkfOkaigx-OdR__kAgYTw

Original Article Link: https://spj.science.org/doi/10.34133/research.1145