Deep learning of structural MRI predicts fluid, crystallized, and general intelligence

Area of Science:

• Neuroimaging
• Cognitive Neuroscience
• Artificial Intelligence

Background:

• Population-level correlations exist between brain structure and intelligence.
• Individual variability in intelligence is not fully explained by population-level associations.
• Existing predictive studies primarily focus on fluid intelligence, with limited research on crystallized or general intelligence.

Purpose of the Study:

• To investigate the potential of deep learning on structural MRI (sMRI) to predict individual intelligence quotients (verbal, performance, and full-scale).
• To explore the prediction of crystallized and general intelligence using sMRI data.
• To assess the relationship between deep learning model complexity and prediction accuracy.

Main Methods:

• Utilized T1-weighted sMRI data from 850 healthy and autistic subjects (ages 6-64).
• Conducted 432 experiments using various input channels, six deep learning models (including 2D and 3D CNNs), and two outcome settings.
• Employed GradCAM for model interpretation to understand brain regions involved in prediction.

Main Results:

• Demonstrated a statistically significant association between sMRI and intelligence prediction (Pearson correlation > 0.21, p < 0.001).
• Found that increased deep learning model complexity did not consistently improve prediction accuracy.
• Model interpretations supported the Parieto-Frontal Integration Theory (P-FIT), highlighting the role of occipital, temporal, parietal, and frontal lobes.

Conclusions:

• T1-weighted sMRI holds significant potential for predicting individual intelligence.
• Deep learning models can predict various intelligence quotients, offering insights into crystallized and general intelligence.
• Findings support the P-FIT theory, suggesting intelligence arises from integrated brain region activity.