The role of machine learning, deep learning, and MRI findings in the classification of pediatric posterior fossa tumors

Abstract

Pediatric posterior fossa tumors represent a major subset of childhood central nervous system neoplasms; however, overlapping MRI features often hinder accurate non-invasive characterization. This study aimed to evaluate machine learning (ML) and deep learning (DL) models for classifying these tumors using MRI-derived radiomic features. Methods This retrospective study analyzed MRI data from 63 pediatric patients with confirmed posterior fossa tumors, including 21 medulloblastoma (MB), 20 pilocytic astrocytoma (PA), 11 ependymoma (EP), and 11 diffuse midline glioma (DMG) cases. T2-weighted imaging, diffusion-weighted imaging, and apparent diffusion coefficient sequences showing the best single-model performance were used to construct gradient boosting machine (GBM), decision tree (DT), and random forest (RF) models and their ensemble combinations. Model performance was evaluated using standard classification metrics. A ResNet101V2-based DL model was developed using multiple MRI sequences. ML models were validated using fivefold cross-validation, whereas the DL model was trained using a 67/33 train–test split with data augmentation. Results The RF + GBM ensemble achieved the highest ML performance, with an overall accuracy of 78%, and showed the strongest classification for MB and PA, whereas EP and DMG remained difficult to distinguish. The DL model demonstrated high performance on T1-weighted imaging and contrast-enhanced T1-weighted imaging, achieving accuracies of 98% and 96%, respectively, but showed lower performance on diffusion-based sequences. Conclusion ML and DL approaches improve MRI-based classification of pediatric posterior fossa tumors; however, accurate differentiation of EP and DMG remains challenging. These findings support the potential of AI-driven methods as clinically relevant, non-invasive decision-support tools in pediatric neuro-oncology.