Abstract
BACKGROUND: Accurate radiographic diagnosis of pediatric brain tumors (PBTs) that originate in the brainstem and posterior fossa, including medulloblastoma (MB), pilocytic astrocytoma (PA), ependymoma (EPN), atypical teratoid/rhabdoid tumor (ATRT), and diffuse intrinsic pontine glioma (DIPG), is crucial for optimizing surgical approaches and enhancing neoadjuvant therapies. Existing research on the radiographic differential diagnosis of posterior fossa tumors has limitations, including small sample sizes, lack of inclusion of certain histologies especially rarer tumors such as ATRTs, and incomplete analysis of the whole tumor, including peritumoral edema. In this study, we aimed to perform a comprehensive analysis using radiomics and machine learning to differentiate among the common posterior fossa and brainstem tumors. METHODS: We employed 927 radiomic features extracted from whole tumor regions within treatment-naïve, standard multiparametric MRI sequences (pre-/post-contrast T1-weighted, T2-weighted, FLAIR) of 264 patients (106 MBs, 78 PAs, 28 EPNs, 27 ATRTs, and 25 DIPGs), collected from the Children’s Brain Tumor Network (CBTN). We adopted a one-versus-rest classification strategy, employing Support Vector Machines combined with the Least Absolute Shrinkage and Selection Operator (LASSO) for feature selection, and implementing nested cross-validation for robustness. RESULTS: The performances of the classifiers were evaluated using the Area Under the Receiver Operating Characteristic Curve, yielding values of 0.84 for MBs, 0.84 for PAs, 0.70 for EPNs, 0.75 for ATRTs, and 0.71 for DIPGs. CONCLUSIONS: Our method effectively differentiates between various tumor types in the posterior fossa and brainstem, paving the path towards the development of comprehensive diagnostic and prognostic AI tools for pre-treatment histological diagnosis of these tumors. These AI tools can lead to more tailored, risk-adjusted treatments for PBTs, reducing morbidities and improving patient outcomes. Based on our promising initial results, we will expand our dataset to include more samples and incorporate rarer tumor types, as well as piloting in different molecular subclassifications.