Application of Radiomics in Predicting the Prognosis of Medulloblastoma in Children.

Abstract

Background and purpose: Medulloblastoma (MB) represents the predominant intracranial neoplasm observed in pediatric populations, characterized by a five-year survival rate ranging from 60% to 80%. Anticipating the prognostic outcome of medulloblastoma in children prior to surgical intervention holds paramount significance for informing treatment modalities effectively. Radiomics has emerged as a pervasive tool in both prognostic anticipation and therapeutic management across diverse tumor spectra. This study aims to develop a radiomics-based prediction model for the prognosis of children with MB and to validate the contribution of radiomic features in predicting the prognosis of MB when combined with clinical features.

Materials and methods: Patients diagnosed with medulloblastoma at our hospital from December 2012 to March 2022 were randomly divided into a training cohort (n = 40) and a test cohort (n = 41). Regions of interest (ROIs) were manually drawn on T1-weighted images (T1WI) along the boundary of the tumor, and radiomic features were extracted. Radiomic features related to survival prognosis were selected and used to construct a radiomics model. The patients were classified into two different risk stratifications according to the Risk-score calculated from the radiomics model. The log-rank test was used to test the difference in survival between the two stratifications to verify the classification value of the radiomics model. Clinical features related to the prognosis were used to construct a clinical model or clinical-radiomics model together with the radiomic features. Then, the clinical model, radiomics model, and clinical-radiomics model were compared to validate the improvement of radiomics in predicting the prognosis of medulloblastoma. The performance of the three models was evaluated with the C-index and the time-dependent AUC. Overall survival (OS) was defined as the time from receiving the operation to death or last follow-up.

Results: A total of 81 children were included in this study. A total of five prognostic radiomic features were selected. The radiomics model could discriminate different risk hierarchies with good performance power in the training and testing datasets (training set p= 0.0009; test set p = 0.0286). Six clinical features associated with prognosis (duration of disease, risk hierarchy, dissemination, radiology, chemotherapy, and last postoperative white blood cell (WBC) level in CSF) were selected. The radiomic-clinical molecular features had better predictive value for OS (C-index = 0.860; Brier score: 0.087) than the radiomic features (C-index = 0.762; Brier score: 0.073) or clinical molecular characteristics (C-index = 0.806; Brier score: 0.092).

Conclusions: Radiomic features based on T1-weighted imaging have predictive value for pediatric medulloblastoma. Radiomics has incremental value in predicting the prognosis of MB, and clinical-radiomics models have a better predictive effect than clinical models.