Multi-Modal Learning for Predicting the Genotype of Glioma.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Oncology

Background:

• Isocitrate dehydrogenase (IDH) gene mutation is a critical biomarker for glioma diagnosis and prognosis.
• Predicting glioma genotype can be enhanced by integrating diverse data types, including imaging and brain network features.
• Current deep learning models struggle with non-Euclidean data like geometric and network features.

Purpose of the Study:

• To develop a multi-modal learning framework for predicting glioma genotype.
• To integrate focal tumor imaging, geometric features, and brain network data.
• To improve the accuracy of isocitrate dehydrogenase (IDH) mutation prediction.

Main Methods:

• A multi-modal learning framework with separate encoders for image, geometry, and brain network features.
• A self-supervised method to generate brain networks from anatomical MRI, addressing diffusion MRI limitations.
• A hierarchical attention module for brain network feature extraction.
• A bi-level multi-modal contrastive loss for feature alignment and domain gap mitigation.
• A weighted population graph for integrating multi-modal features for genotype prediction.

Main Results:

• The proposed model significantly outperforms baseline deep learning models in predicting glioma genotype.
• Ablation experiments confirm the effectiveness of individual framework components.
• Visualized interpretations align with existing clinical knowledge, validating the model's insights.

Conclusions:

• The developed multi-modal learning framework offers a novel and effective approach for glioma genotype prediction.
• Integrating diverse data modalities, including self-supervised brain networks, enhances predictive accuracy.
• This framework has the potential to improve glioma diagnosis and prognosis through precise genotype identification.