IMG-43. Non-Invasive Prediction of Malignant Transformation in Grade 2 IDH-Mutant Gliomas Using Radiomics and Machine Learning

Grade 2 IDH mutant gliomas are known to undergo malignant transformation (MT). Early prediction of this transformation remains a significant clinical challenge. Radiomics, the extraction of high-dimensional imaging features, has emerged as a promising tool for non-invasive tumor characterization. Our study looked to implement quantitative radiomics analyses to identify textural information from clinical imaging that might enhance the prediction of MT in grade 2 IDH-mutant gliomas. We conducted a retrospective analysis of institutional (UCLA) and public (UCSF-PDGM) datasets with histologically confirmed IDH-mutant gliomas. Institutional dataset included patients underwent serial magnetic resonance imaging (MRI) and had pathologic clinical progression and/or transformation with at least one repeat biopsy/surgery. Radiomic features were extracted from T1-subtraction maps, T2-weighted images, and apparent diffusion coefficient (ADC) maps from diffusion MRI scans. The T2-hyperintense lesion was segmented, excluding necrosis, using NS-HGlio (Neosoma Inc, Groton, MA). UCLA patients were categorized based on the presence or absence of MT, defined as progression from WHO grade 2 to grade 3 or 4 by histology at repeat biopsy/surgery. Machine learning models, including random forest and support vector machines, were trained on 75% UCSF-PDGM dataset to classify grade 2 versus grade 3/4 tumors, validated on the remaining 25%. The model was used to generate a radiomics risk-score to predict MT in the UCLA dataset, evaluated using the ROC analysis. Model achieved AUC of 0.814 (CI: 0.73--0.90) on the UCSF training set (n = 77). On the UCSF test set (n = 26), sensitivity was 0.93, specificity 0.92, with AUC of 0.92 (CI: 0.80–1.00). Evaluation on the independent dataset of UCLA patients at the second surgical timepoint (n = 59) demonstrated a sensitivity of 0.69, specificity of 0.76, and AUC of 0.73 (CI: 0.61–0.84). Radiomics may enable non-invasive prediction of MT in IDH-mutant gliomas, outperforming conventional clinical models.