Objectives: Conventional MRI may not be ideal for predicting cervical spondylotic myelopathy (CSM) prognosis. In this study, we used radiomics in predicting postoperative recovery in CSM. We aimed to develop and validate radiomic feature-based extra trees models.
Methods: There were 151 patients with CSM who underwent preoperative T2-/ T2*-weighted imaging (WI) and surgery. They were divided into good/poor outcome groups based on the recovery rate. Datasets from multiple scanners were randomised into training and internal validation sets, while the dataset from an independent scanner was used for external validation. Radiomic features were extracted from the transverse spinal cord at the maximum compressed level. Threshold selection algorithm, collinearity removal, and tree-based feature selection were applied sequentially in the training set to obtain the optimal radiomic features. The classification of intramedullary increased signal on T2/T2*WI and compression ratio of the spinal cord on T2*WI were selected as the conventional MRI features. Clinical features were age, preoperative mJOA, and symptom duration. Four models were constructed: radiological, radiomic, clinical-radiological, and clinical-radiomic. An AUC significantly > 0.5 was considered meaningful predictive performance based on the DeLong test. The mean decrease in impurity was used to measure feature importance. p < 0.05 was considered statistically significant.
Results: On internal and external validations, AUCs of the radiomic and clinical-radiomic models, and radiological and clinical-radiological models ranged from 0.71 to 0.81 (significantly > 0.5) and 0.40 to 0.55, respectively. Wavelet-LL first-order variance was the most important feature in the radiomic model.
Conclusion: Radiomic features, especially wavelet-LL first-order variance, contribute to meaningful predictive models for CSM prognosis.
Key points: • Conventional MRI features may not be ideal in predicting prognosis. • Radiomics provides greater predictive efficiency in the recovery from cervical spondylotic myelopathy.
Keywords: Machine learning; Magnetic resonance imaging; Radiomics; Spinal cord diseases.
© 2022. The Author(s), under exclusive licence to European Society of Radiology.