108 Interpretable Radiomics-Based Prediction of IDH Status in Glioma

Abstract Introduction Glioblastoma (GBM) is the most frequent primary brain tumour. The 2021 WHO CNS Tumour Classification highlights the importance of IDH mutation status for accurate diagnosis and treatment. However, biopsies are invasive and may not fully capture tumour heterogeneity in GBM. We explored if non-invasive MRI-based radiomic analysis can predict the IDH mutation status of Glioma. Method Utilising the open-access UCSF and a local cohort (STORM_GLIO), a multi-institutional dataset (n = 376, Grade 4) was employed to classify patients with IDH-wildtype and IDH-mutated Glioma. IDH mutation status was predicted via traditional machine learning models using up to 10 standardised radiomic features extracted from Gross Tumour Volume (GTV) region using an IBSI standardised radiomic analysis software (spaarc-radiomics.io). Results The models were trained and validated on a discovery cohort comprising 80% of the dataset, utilizing an XGBoost classifier with LASSO-selected features. Our clinical-radiomic model achieved an AUC of 0.97 on the internal validation set. Additionally, the model yielded an AUC of 0.93 on the independent external validation cohort (20% of the data). Model interpretability, assessed via SHAP analysis, reported ivh_i90 (MRI T2-dependent) and dzm_sdhge_3d (MRI FLAIR-dependent) as the most impactful radiomic predictors of IDH status, specifically capturing high-intensity volumes and high-grey level zones within enhancing tumour tissues. Along with these imaging biomarkers, Age ranked among the top three most impactful features in the external validation cohort. Conclusions These findings highlight the utility of standardised radiomic features in predicting IDH status in Glioma. Clinical integration of imaging radiomic studies could allow robust patient stratification based on IDH mutation status and the potential addition of a diagnostic imaging biomarker to diagnostic pathway.