187 Decoding Diffuse Gliomas: A Radiomics-Driven Machine Learning Strategy for Predicting Histopathological and Molecular Profiles

INTRODUCTION: Accurate preoperative prediction of histopathological and molecular profiles of diffuse gliomas using MRI images could be essential for personalized treatment planning. METHODS: The TCGA-GBM and TCGA-LGG datasets were used for analysis. Radiomic features were extracted using PyRadiomics from T1, T1 with contrast, T2, and FLAIR MRI sequences. LASSO regression was employed for feature selection. The TabPFN algorithm was used for model development, and a 10-repeat stratified 5-fold-cross-validation approach was used for model evaluation. A framework for a fully automatic and user-friendly web application was proposed to make these models accessible to clinicians and researchers worldwide. RESULTS: The models achieved high performance across all prediction tasks. For tumor grade and histology, the mean AUROCs were 0.921 and 0.901, respectively. The IDH mutation status model showed the highest performance with an AUROC of 0.970. The models for other molecular markers also demonstrated strong discriminatory abilities, with mean AUROCs of 0.898 for 1p19q codeletion, 0.938 for TERT promoter status, 0.911 for EGFR amplification, 0.905 for concomitant gain of chromosome 7 and loss of chromosome 10, 0.887 for homozygous CDKN2A/B deletion, 0.910 for ATRX mutation status, and 0.885 for MGMT promoter status. CONCLUSIONS: Radiomics-based machine learning models show promise for noninvasive prediction of the histopathological and molecular profiles of diffuse gliomas based on preoperative MRI images. These models could potentially aid in glioma characterization and guide personalized treatment strategies. The proposed web application framework would make these models accessible to clinicians and researchers worldwide, facilitating their integration into clinical practice.