Radiomics and machine learning for Pfirrmann grade classification of intervertebral discs in lumbar MRI

Abstract

Intervertebral disc degeneration (IDD) is a leading cause of chronic low back pain, yet its clinical grading with the Pfirrmann scale is subjective and prone to variability. This study evaluates a radiomics-based machine learning framework for automatic classification of Pfirrmann grades from lumbar T2-weighted MRI. Radiomic features of disc shape, intensity, and texture were extracted under IBSI guidelines and classified using six machine learning models with patient-level cross-validation, Bayesian hyperparameter optimization, and probability calibration. Gradient Boosting achieved the best overall performance, with a mean AUC of 0.87 in multiclass classification and 0.94 in binary classification, the latter improving sensitivity to advanced degeneration and alleviating class imbalance. SHapley Additive exPlanations (SHAP) identified texture descriptors and shape sphericity as key predictors, with feature patterns aligning with physiologic degeneration stages. These results demonstrate that radiomics combined with machine learning provides accurate and interpretable grading of disc degeneration, offering a reproducible and clinically meaningful alternative to subjective visual assessment.