Revealing Distinctive Insights: Machine Learning-Enhanced Ensembled MRI Sequences for Pediatric Posterior Fossa Tumor Classification

Abstract Pediatric posterior fossa (PF) tumors pose a significant concern in pediatric oncology because of their high prevalence in the pediatric population. Magnetic Resonance Imaging (MRI) has emerged as a practical and non-invasive imaging modality for detecting and classifying PF malignancies. For the differentiation of MB, the LGBM model, using the feature combination of T2, FLAIR, DWI, and ADC, demonstrated the highest performance. It achieved an AUROC of 0.938, accuracy of 0.778, specificity of 0.916, sensitivity of 0.645, and F1 score of 0.664. In the classification of EP, the XGB model, which utilizes the feature combination of T2 and DWI, demonstrated the highest performance with a sensitivity of 1. With respect to the distinction of PA, the RF model, using the feature combination of T2, T1CE, DWI, and ADC, exhibited the highest performance. For the classification of BG, the RF model, using the feature combination of T2, DWI, and ADC, exhibited the highest performance. This comprehensive approach has notably enhanced our comprehension of pediatric PF tumor classification and carries substantial potential for the advancement of diagnostic tools and the refinement of ML models.