Work

Quantitative Grading of Disc Health and Degeneration from Routine MRI: AI-Based Prediction of T2 Relaxometry Values using Radiomics and a proposal for New Grading

Year: 2026
Type: preprint
Abstract: Purpose Identifying Disc degeneration (DD) before it is obvious on routine Magnetic Resonance Imaging (MRI) is the focus of spine research. Quantitative T2 relaxometry of the nucleus pulposus (T2r) offers objective biochemical evaluation of DD but has disadvantages of high costs, long scan times, and limited availability. We evaluated whether radiomic features derived from standard sagittal T2-weighted(T2W) MRI could accurately predict degeneration status and estimate continuous T2r values, thereby making T2r values available from normal T2W images. Methods We analyzed 621 lumbar intervertebral discs from 160 individuals. A data-derived T2r degeneration threshold was established using a reference cohort of 334 volunteer discs. Discs were partitioned at the patient level into training (353), validation (154), and external validation test set (114 discs). 111 radiomic features were extracted using 3D Slicer. A quantile-transformed logistic regression model was trained to predict binary degeneration (T2r ≤ 82.2 ms), while a Support Vector Regression (SVR) model estimated continuous T2r values. Results The binary radiomics model demonstrated good prediction with an AUC of 0.928 (95% CI: 0.87–0.98) and accuracy of 86.0% on the held-out test set. Importantly, the model correctly identified 70.8% of biochemically altered discs within Pfirmann I and II discs. For continuous modelling, SVR achieved a strong Spearman correlation (ρ = 0.79) with actual T2r values, yielding a low median absolute deviation of 12.2 ms. Conclusion T2r radiomics provides a promising surrogate for advanced quantitative mapping and provides universal access to biochemical disc profiling without need for specialised software, increased time, and costs.
Institution: Ganga Hospital
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7236342989
Cites: 23
Cited by:
Related to: 10