Clinical-MRI Radiomics Enables the Prediction of Preoperative Cerebral Spinal Fluid Dissemination in Children With Medulloblastoma 

Abstract

Background: Medulloblastoma (MB) is the most common pediatric embryonal tumor. Accurate identification of cerebral spinal fluid (CSF) dissemination is important in prognosis prediction. Both MRI of the central nervous system (CNS) and CSF cytology will appear false positive and negative. Our objective was to investigate the added value of enhanced T1 weighted images-based radiomics features to clinical characteristics in predicting preoperative CSF dissemination for children with MB.Materials and Methods: This retrospective study included 84 children with histopathologically confirmed MB (60 children in the training cohort and 24 children in the validation cohort). The children with normal head and spine magnetic resonance images (MRI) and no subsequent dissemination in one year were diagnosed as non-CSF dissemination. The CSF dissemination was manifested as intra-cranial or -spinal nodular enhanced lesions. Clinical features and conventional MRI features were collected and evaluated. A total of 385 radiomics features were extracted from enhanced T1 weighted images. Minimum redundancy, maximum correlation and least absolute shrinkage and selection operator were performed to select the features with the best performance in predicting preoperative CSF dissemination. A combined clinical-MRI radiomics prediction model was developed using multivariable logistic regression. Receiver operating curve analysis was used to validate the predictive performance. Nomogram and decision curve analysis (DCA) were developed to evaluate the clinical utility of combined model.Results: Thirty-one children were confirmed to have preoperative CSF dissemination. One clinical and nine radiomics features were selected for predicting preoperative CSF dissemination. The combined model incorporating clinical and radiomics features had best predictive performance both in the training and validation cohorts. In the validation cohort, a threshold of 0.311 yielded an area under the curve of 0.87, a sensitivity of 77.8%, a specificity of 86.7%, and an accuracy of 83.3%. The clinical utility of the model was confirmed by a clinical-MRI radiomics nomogram and DCA. Conclusions: The combined model incorporating clinical, conventional MRI and radiomics features could be applied to predict preoperative CSF dissemination for children with MB as a noninvasive biomarker, which could aid in risk evaluation.