Magnetic Resonance Imaging-Based Radiomics for Quantitative Assessment of Cervical Disc Degeneration: A Dual-Cohort Study

Objective: Cervical disc degeneration (CDD) is commonly evaluated using magnetic resonance imaging (MRI)-based Pfirrmann grading, but this visual semiquantitative system remains observer-dependent and may not fully capture intradiscal heterogeneity. This study aimed to develop and externally validate a low-dimensional, interpretable MRI-based radiomics model for identifying Pfirrmann-defined moderate-to-severe CDD using routine midsagittal T2-weighted imaging.Methods: This retrospective dual-cohort study included 374 patients with 1,870 discs in the internal cohort and 66 patients with 330 clinically graded discs in the external cohort. Quantitative features reflecting signal distribution, texture-based spatial gray-level relationships, and morphology were extracted from manually delineated disc regions of interest on midsagittal T2-weighted images. A multivariable logistic regression model was constructed, and internal validation was performed using patient-level 5-fold cross-validation. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and patient-level bootstrap resampling. Calibration was assessed using calibration curves and the Hosmer-Lemeshow test, and imaging-based net benefit was evaluated using decision curve analysis (DCA).Results: Most selected features showed good reproducibility, whereas 2-dimensional (2D) sphericity was less reproducible. In patient-level internal cross-validation, the combined model achieved the highest AUC (0.821; 95% CI, 0.797–0.846), exceeding that of the mean-intensity baseline model (0.677; 95% CI, 0.641–0.713; patient-level bootstrap p<0.001). In the independent external cohort, the combined model achieved an AUC of 0.832 (95% CI, 0.786–0.876), with 76.7% accuracy, 69.2% sensitivity, and 83.6% specificity. In the ablation analysis, removal of 2D sphericity reduced the external AUC to 0.807 and specificity to 63.2% but increased sensitivity to 83.0%; the AUC difference was not statistically significant (ΔAUC=0.025; 95% CI, -0.009 to 0.058; p=0.152). Patient-cluster sensitivity analyses yielded generally consistent adjusted associations after accounting for within-patient clustering. Calibration showed no evidence of poor fit (Hosmer-Lemeshow χ2=10.103, df=8, p=0.258), and DCA indicated potential imaging-based net benefit across the evaluated threshold range.Conclusion: A low-dimensional, interpretable radiomics model based on routine midsagittal T2-weighted MRI showed stable performance for identifying Pfirrmann-defined moderate-to-severe CDD. The model may serve as a quantitative adjunct for standardized imaging-based degeneration assessment, but further validation against clinical outcomes and quantitative biochemical MRI markers is required.