Document Type : Original Research

Authors

• Somayeh Farahani ^{1, 2, 3}
• Marjaneh Hejazi ^{1, 4}
• Antonio Di Ieva ³
• Emad Fatemizadeh ⁵
• Ahmad Pour-Rashidi ¹
• Sidong Liu ^{2, 3}

¹ Department of Medical Physics and Biomedical Engineering, School of Medicine, Tehran University of Medical Sciences, Tehran, Iran

² Centre for Health Informatics, Australian Institute of Health Innovation, Macquarie University, Sydney, NSW, Australia

³ Computational NeuroSurgery (CNS) Lab, Faculty of Medicine, Health and Human Sciences, Macquarie Medical School, Macquarie University, Sydney, NSW, Australia

⁴ Research Center for Molecular and Cellular Imaging, Bio-Optical Imaging Group, Tehran University of Medical Sciences, Tehran, Iran

⁵ Department of Electrical Engineering, Sharif University of Technology, Tehran, Iran

Abstract

Background: Glioma management depends on molecular profiling, particularly Isocitrate Dehydrogenase (IDH) status, but requires invasive biopsy. Radiomics applied to Magnetic Resonance Imaging (MRI) offers a non-invasive alternative, though translation is hindered by variability in segmentation, limited external validation, and lack of interpretability.
Objective: We aimed to develop and validate an interpretable radiomics pipeline for non-invasive IDH classification across MRI data.
Material and Methods: In this retrospective study, we assembled 1117 preoperative MRI scans from three public cohorts: The Cancer Genome Atlas (TCGA), the University of California San Francisco Preoperative Diffuse Glioma MRI (UCSF-PDGM), and the Erasmus Glioma Database (EGD). Tumor subregions were segmented using BrainSegFounder model. From the segmented regions, we extracted approximately 1900 PyRadiomics features per patient. Boruta selection was applied, followed by classification with k-nearest neighbors (kNN), Light Gradient Boosting Machine (LightGBM), and Logistic Regression (LR). Model interpretability was evaluated using Shapley Additive Explanations (SHAP).
Results: All classifiers achieved strong internal performance. On the external EGD test set, kNN provided the most balanced generalization (area under the receiver operating characteristic curve [AUC]=0.94, Matthews correlation coefficient [MCC]=0.74), whereas LR maximized sensitivity (0.93). On the TCGA test set, LR provided the best overall balance (AUC=0.91, MCC=0.68), and LightGBM achieved the highest specificity (0.94). SHAP analysis highlighted texture features derived from post-contrast T1-weighted images as key predictors. 
Conclusion: Our pipeline achieves high IDH classification performance with full automation, cross-site generalizability, and transparency. These findings support its potential for clinical application in non-invasive glioma genotyping.

Keywords

• Glioma
• Machine Learning
• Image Segmentation
• Isocitrate Dehydrogenase
• Radiomics