MRI-based severity classification of central canal stenosis using radiomics features and machine learning

Accurate assessment of lumbar degenerative disease is important for treatment planning and surgical decision-making. In this study, we investigated whether radiomics features extracted from lumbar MRI, combined with machine learning models, could be used to classify the severity of lumbar spinal stenosis.

We analyzed 1,500 sagittal T2/STIR-weighted lumbar MRI images representing three categories of central canal stenosis severity: normal/mild, moderate, and severe. Two experienced radiologists manually segmented the sagittal slices, and radiomics features were extracted from each region of interest. These features were used to train and evaluate four machine learning models: random forest (RF), adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), and an artificial neural network (ANN), for stenosis severity classification.

All models demonstrated high classification performance, with AdaBoost achieving the highest overall effectiveness. Specifically, AdaBoost reached an accuracy of 80% (95% CI: 75.47-84.53%) and a precision of 81% (95% CI: 76.56- 85.44%). The multi-class Area Under the Curve (AUC) values were 93.60% (95% CI: 90.83- 96.37%) for AdaBoost, 92.80% (95% CI: 89.88-95.72%) for XGBoost, 92.60% (95% CI: 89.64- 95.56%) for RF, and 90.70% (95% CI: 87.41-93.99%) for ANN. While AdaBoost led in discrimination, the ANN model achieved the highest specificity at 90% (95% CI: 86.60-93.40%).

The results of this study support the feasibility of using radiomics-based machine learning for severity classification of central canal stenosis on lumbar MRI. Although the findings are encouraging, the current workflow relies on manual segmentation and should be regarded as an exploratory proof of concept. Further work is needed to evaluate generalizability and to integrate automated methods suitable for clinical use.