Abstract Details

Title Robust Machine Learning Techniques Accurately Predict Genotype Using Magnetic Resonance Radiomic Data: A Multicentric Pooled Analysis of 263 Glioma Patients

Topic Neuro-oncology

Presentation(s) P4 - Poster Session 4 (5:00 PM-6:00 PM)

Poster/Presentation
Number 6-001

Objective To explore the potential of radiomics as a non-invasive diagnostic tool to predict 1p/19q co-deletion status, IDH mutant genotype, and MGMT promoter methylation.

Background Gliomas are brain tumors with varying prognoses and treatment strategies significantly influenced by genetic alterations. These can provide insights into drug development by acting as novel drug targets. Among these, the 1p/19q co-deletion, isocitrate dehydrogenase (IDH) mutation status, and O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation are a few important biomarkers. Radiomics refers to the extraction of quantitative features from radiographic images that can be analyzed to provide information not possible with the human eye alone.

Design/Methods We used two open-source datasets—the updated LGG-1p19qDeletion (2020) and ReMIND (2024) datasets, with a cumulative pool of 263 patients. After preprocessing the images, we extracted 2246 radiomic features per patient using the PyRadiomics library. After using ensemble feature selection techniques, we reduced the number of features to less than 50 per model. We trained logistic regression, Support Vector Machine (SVM), random forest (RF), and eXtreme Gradient Boosting (XGBoost) models to predict tumor genotype and assessed their performance. Predictions were explained using local interpretable model-agnostic explanations (LIME).

Results The RF model showed the highest accuracy (79%) and Dice score (0.778) predicting 1p/19q co-deletion status. For IDH mutation status, RF was the best-performing model, with an accuracy of 84.6% and a Dice score of 0.819. Finally, the SVM model was the highest-performing model to predict MGMT promoter methylation status with an accuracy of 81.5% and a Dice score of 0.80. Precision, recall, and specificity were above 75% for all three models. LIME analysis revealed that gray-level dependence features best-explained predictions for all three models.

Conclusions Our results suggest that radiomic features alone exhibit robust performance in predicting genomic features. These models can be used as alternate non-invasive diagnostic tools to guide therapeutic regimen selection after clinical validation.

Authors/Disclosures
Vinay Suresh, MBBS PRESENTER	Dr. Suresh has nothing to disclose.
Suhrud Panchawagh, MBBS	Dr. Panchawagh has nothing to disclose.

��ɫ��

��ɫ��