A Systematic Evaluation of MRI Normalization for Multi-Site Radiomics-Based Disc Degeneration Classification

ABSTRACT Objective Automated grading of intervertebral disc degeneration from MRI is clinically important but challenging due to inter-rater variability in the Pfirrmann classification and variations in signal intensity and contrast across MRI scanners. Automated approaches may require image normalization to harmonize data from different scanners and ensure robust predictions. Although radiomics-based disease classification offers advantages in explainability and efficiency over deep learning, the effect of image normalization on feature stability and classification performance remains unclear. This study systematically investigated eight MRI normalization pipelines for disc classification and introduced a normalization-robust, interpretable grading tool with performance comparable to human raters. Methods 270 T2-weighted spine MRIs from different scanners and field strengths were used for training, testing, and validation. Repeat scans from nine subjects enabled assessment of radiomic feature reproducibility. Features were extracted from whole-disc volumes and evaluated across eight normalization pipelines and an unnormalized control. Feature selection prioritized both classification relevance and scan-rescan reproducibility. XGBoost was used for classification. Results Normalization improved feature reproducibility, but classification performance remained statistically equivalent across all pipelines, indicating robustness to scanner variations. The XGBoost model showed strong agreement with expert grading (Concordance Correlation Coefficient = 0.77), comparable to reported inter-rater agreements. Key features reflected nucleus pulposus desiccation and structural disorganization, supporting biological interpretability. A practical normalization decision framework is presented. Conclusion Radiomics-based disc degeneration grading is robust to MRI normalization strategy, supporting deployment across different scanners. Its reproducibility, interpretability, and multi-site validation position it as a practical alternative to subjective grading and black-box deep learning.