Machine learning-based MRI radiomics identifies patients with degenerative cervical myelopathy and predicts baseline function.

purposeThe diagnosis of Degenerative cervical myelopathy (DCM) relies on clinical evaluation and conventional MRI, yet early symptoms are subtle and non-specific, creating diagnostic uncertainty. High rates of asymptomatic cervical spinal cord compression in the older adults further complicate clinical decisions regarding surgical intervention. MRI-based radiomic analysis quantifies microstructural spinal cord pathology through texture analysis, and is a potential objective imaging biomarker. This study aimed to assess whether machine learning models using MRI radiomics can accurately classify DCM and predict disease severity. We also aim to evaluate the impact of imaging resolution on diagnostic performance.

methodsWe included 79 patients with clinically diagnosed DCM undergoing surgical evaluation and 51 healthy controls. Participants underwent high-resolution 3D T2-weighted MRI. Images were processed using automated spinal cord segmentation, and radiomic features were extracted using the Pyradiomics software package and filtered by correlation, variance, ANOVA, and mutual information analyses. Machine learning algorithms (XGBoost, CatBoost, LightGBM, Random Forest, SVM) underwent hyperparameter optimization with Optuna and were evaluated using 5-fold cross-validation. For DCM versus healthy control classification, model performance was additionally assessed in an internal held-out test set; however, external validation in an independent dataset was not performed. Disease severity prediction was evaluated using 5-fold cross-validation only because of the limited sample size.

resultsMachine learning using MRI radiomics discriminated DCM from healthy controls with high accuracy in the internal held-out test set (AUROC = 0.93, accuracy = 0.88, F1-score = 0.90, MCC = 0.76, Precision = 93.3%, sensitivity = 87.5%, specificity = 90%). Radiomic features from high-resolution 3D MRI showed superior diagnostic performance compared to standard MRI and conventional morphometry. The models effectively classified disease severity (macro-average AUROC = 0.855 ± 0.062), significantly outperforming traditional imaging measures.

conclusionMRI-based radiomics combined with machine learning demonstrates robust accuracy in predicting DCM status and estimating disease severity, independent of clinical or demographic data. These findings underscore the potential of MRI radiomics as an objective imaging biomarker to screen for DCM. However, external validation was not performed in this study, and future validation in larger, multi-vendor independent datasets is needed before broader clinical application.