GlioVision: A Multi-Modal MRI Framework for Non-Invasive Glioma Molecular Biomarkers Prediction.

Area of Science:

• Neuro-oncology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Gliomas are aggressive primary brain tumors requiring molecular biomarker prediction for clinical decisions.
• Current methods involve invasive surgical tumor analysis, risking complications and sampling bias.
• Existing non-invasive deep learning models are often resource-intensive and trained on limited data.

Purpose of the Study:

• To present GlioVision, a framework for non-invasive prediction of four major glioma molecular biomarkers.
• To enable real-time, resource-efficient, and accurate molecularly defined glioma diagnosis.
• To support clinical decision-making using the WHO 2021 classification guidelines.

Main Methods:

• GlioVision utilizes the MONAI library to process multimodal glioma MRI and molecular data.
• The core architecture, SCRU-DenseNet, incorporates attention gates and adaptive contrast processing for heterogeneous datasets.
• Confidence-Filtered Predictive Manifold (CFPM) manages prediction uncertainty; Differential Training Integrity Assessment (DTI-A) analyzes data privacy.

Main Results:

• GlioVision achieved high prediction AUCs: IDH (0.94), 1p/19q co-deletion (0.87), MGMT methylation (0.86), and WHO grades (0.92).
• The framework was trained and validated on the largest multi-cohort datasets to date.
• Results demonstrate strong performance supporting molecularly defined glioma diagnosis.

Conclusions:

• GlioVision offers a robust, non-invasive framework for predicting critical glioma molecular biomarkers.
• The framework advances AI in neuro-oncology, improving diagnostic accuracy and efficiency.
• This work contributes to the codebase, model release, and data privacy considerations in MRI analysis.