Multi-parametric MRI-based radiomics to differentiate medulloblastoma from ependymoma in children

Objective To establish and evaluate a radiomics classification model based on multi-parameter MRI radiomics for distinguishing medulloblastoma (MB) form ependymoma (EP) in children. Materials and methods Retrospectively collected MRI images from 113 MB patients and 31 EP patients. Univariate analysis was performed to identify MRI features significant for distinguishing MB from EP. Radiomics features were extracted from MRI sequences including ADC, T2WI and DCE-T1WI. The least absolute shrinkage and selection operator (LASSO) regression analysis was used to select radiomics features for MB and EP, and the radiomics score (Rad-score) was calculated. A combined model integrating MRI findings and radiomics scores was established, and its diagnostic capability was evaluated. Diagnostic performance was compared using the DeLong test between different radiomics models and radiologist diagnoses. Results Among single-sequence models, the ADC sequence model demonstrated the highest diagnostic performance, achieving AUC value of 0.936 and 0.924 in the primary and validation cohorts, respectively. The diagnostic performance surpassed that of the DCE-T1WI sequence model and the T2WI sequence model. Diagnostic performance was further enhanced when the three single-sequence models were combined, achieving AUC value of 0.981 and 0.939 in in the primary and validation cohorts, respectively. The diagnostic efficacy of the combined sequence model outperformed that of radiologists. Conclusion The multiparameter MRI-based radiomics nomogram demonstrates excellent predictive performance for MB and EP, potentially holding significant clinical value for preoperative detection of posterior fossa tumors in pediatric patients.