Predicting Intelligence Profile and Brain Age with Single- and Dual-Channel Cnns: A Study Based on Human Connectome Projects

Many studies formulated brain age biomarkers by applying deep-learning on 3D MR images to predict chronological age, considered a proxy of cognitive functions. Using over 1,700 T1- and T2-weighted images from the Human Connectome Projects (HCP-Young adult and HCP-Aging), we constructed 3D convolutional neural networks to directly predict two major sub-divisions of cognitive functions (fluid vs. crystallized). After competitive performance was obtained for brain age prediction <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(r$</tex> = 0.968, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$MAE$</tex> = 3.327), the more novel intelligence prediction problem was investigated orthogonally with three factors: MRI modality (T1 / T2 / both), prediction target (fluid / crystallized / both), and correction for the “regression to the mean” problem (uncorrected vs. Cole's method). Results showed good performance for both fluid <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(r$</tex> = 0.628) and crystallized intelligence <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(r$</tex> = 0.486). Our findings speak to the promise of the direct approach for predicting cognition and revealed certain advantages of predicting the two intelligences simultaneously over separately.