SpineScan: a deep learning model for lumbar spine MRI annotation and Pfirrmann grading assessment

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Spine Degeneration Research

Background:

• Deep learning advances enable automated Pfirrmann grading for intervertebral disc degeneration (IDD).
• Many current models are proprietary and inaccessible.
• Need for validated, open-source AI tools for IDD assessment.

Purpose of the Study:

• Develop and validate a convolutional neural network (CNN) for automated Pfirrmann grading.
• Utilize a diverse clinical dataset for robust model training.
• Create an open-source web application (SpineScan) for AI-assisted grading.

Main Methods:

• Trained a CNN (YOLOv8x architecture) on 484 lumbar MRI scans from two datasets.
• Model simultaneously detects discs and classifies Pfirrmann grades (I-V) from single slices.
• Evaluated performance using precision, recall, and mean average precision (mAP).

Main Results:

• Achieved predictive accuracy between 0.78-0.82, comparable to expert radiologists.
• Highest performance for Grade IV discs (mAP50=0.872); lower for Grade V (mAP50-95=0.525).
• Overall precision of 0.75 and recall of 0.808.

Conclusions:

• The developed AI model shows strong potential for automated lumbar disc degeneration grading.
• The model's performance is comparable to expert radiologists.
• SpineScan web application supports AI-assisted Pfirrmann grading in clinical settings.