neuro-CaPTk

neuro-CaPTk predicts molecular alterations in gliomas from preoperative magnetic resonance imaging (MRI) using radiomic feature extraction and machine learning for noninvasive radiogenomic characterization.

Key Features:

• Noninvasive Detection: Predicts critical molecular markers including isocitrate dehydrogenase (IDH) mutations, 1p/19q co-deletion, and epidermal growth factor receptor variant III (EGFRvIII) from MRI data.
• Multi-Institutional Data Integration: Developed and evaluated using preoperative MRI data from 473 glioma patients aggregated across Hospital of the University of Pennsylvania (HUP), The Cancer Imaging Archive (TCIA), and Ohio Brain Tumor Study (OBTS).
• Advanced Feature Extraction: Extracts histogram, shape, anatomical, and texture characteristics from delineated tumor subregions.
• Machine Learning Integration: Employs support vector machines (SVM) to integrate extracted features and develop predictive models for IDH, 1p/19q co-deletion, and EGFRvIII status.
• Model Validation and Accuracy: Validated using training-testing splits within individual collections, merged collection analysis, and cross-collection validation, reporting classification accuracies of 86.74% for EGFRvIII, 85.45% for IDH mutations, and 75.15% for 1p/19q co-deletion in the HUP dataset.
• Modular Platform: Implemented as a modular platform to support extension to additional cancer imaging and radio(geno)mic analyses.

Scientific Applications:

• Molecular biomarker prediction: Enables noninvasive inference of tumor genotype (IDH, 1p/19q, EGFRvIII) from MRI-derived radiomic signatures.
• Personalized treatment stratification: Supports tailoring of therapeutic decisions and prognostic assessments based on predicted molecular profiles.

Methodology:

Extracts histogram, shape, anatomical, and texture features from delineated tumor subregions in preoperative MRI; applies support vector machines (SVM) to integrate features and predict IDH, 1p/19q co-deletion, and EGFRvIII status; validates models using training-testing splits within collections, merged collection analysis, and cross-collection validation on data from HUP, TCIA, and OBTS (n=473).