Multi-modal deep learning from imaging genomic data for schizophrenia classification.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Genetics

Background:

• Schizophrenia (SZ) is a complex psychiatric disorder impacting cognition, emotion, and behavior.
• Its exact causes are unknown, but it involves structural and functional brain changes and genetic factors.
• Investigating SZ through multiple data types is crucial for better detection.

Purpose of the Study:

• To develop a deep learning framework for enhanced schizophrenia detection.
• To integrate structural MRI (sMRI), functional MRI (fMRI), and genetic data (SNP).
• To improve classification accuracy by combining multi-modal features.

Main Methods:

• Utilized DenseNet for sMRI morphological features.
• Applied 1D CNN and LRP for fMRI functional connections and SNP relevance.
• Integrated features using Extreme Gradient Boosting (XGBoost) for classification.
• Employed Explainable AI (XAI) for feature interpretation.

Main Results:

• The multi-modal approach achieved 79.01% accuracy in classifying SZ from healthy controls (HC).
• This performance surpassed the accuracy of methods using individual data modalities.
• XAI identified significant functional networks and SNPs contributing to classification.

Conclusions:

• A deep learning framework effectively integrates sMRI, fMRI, and genetic data for improved SZ classification.
• Explainable AI provides insights into key biomarkers for schizophrenia.
• This multi-modal strategy offers a promising avenue for advancing SZ detection and understanding.