Structural brain imaging predicts individual-level task activation maps using deep learning

Abstract

Accurate individual functional mapping of task activations is a potential tool for biomarker discovery and is critically important for clinical care. While structural imaging does not directly map task activation, we hypothesized that structural imaging contains information that can accurately predict variations in task activation between individuals. To this end, we trained a convolutional neural network to use structural imaging (T1-weighted, T2-weighted, and diffusion tensor imaging) to predict 47 different functional MRI task activation volumes across seven task domains. The U-Net model was trained on 654 subjects and then subsequently tested on 122 unrelated subjects. The predicted activation maps correlated more strongly with their actual maps than with the maps of the other test subjects. An ablation study revealed that a model using the shape of the cortex alone or the shape of the subcortical matter alone was sufficient to predict individual-level differences in task activation maps, but a model using the shape of the whole brain resulted in markedly decreased performance. The ablation study also showed that the additional information provided by the T2-weighted and diffusion tensor imaging strengthened the predictions as compared to using the T1-weighted imaging alone. These results indicate that structural imaging contains information that is predictive of inter-subject variability in task activation mapping and cortical folding patterns as well as microstructural features may be a key component to linking brain structure to brain function.