Mapping of the Language Network With Deep Learning

Abstract

Background: Pre-surgical functional localization of eloquent cortex with task-based functional MRI (T-fMRI) is part of the current standard of care prior to resection of brain tumors. Resting state fMRI (RS-fMRI) is an alternative method currently under investigation. Here, we compare group level language localization using T-fMRI vs. RS-fMRI analyzed with 3D deep convolutional neural networks (3DCNN). Methods: We analyzed data obtained in 35 patients with brain tumors that had both language T-fMRI and RS-MRI scans during pre-surgical evaluation. The T-fMRI data were analyzed using conventional techniques. The language associated resting state network was mapped using a 3DCNN previously trained with data acquired in >2,700 normal subjects. Group level results obtained by both methods were evaluated using receiver operator characteristic analysis of probability maps of language associated regions, taking as ground truth meta-analytic maps of language T-fMRI responses generated on the Neurosynth platform. Results: Both fMRI methods localized major components of the language system (areas of Broca and Wernicke). Word-stem completion T-fMRI strongly activated Broca's area but also several task-general areas not specific to language. RS-fMRI provided a more specific representation of the language system. Conclusion: 3DCNN was able to accurately localize the language network. Additionally, 3DCNN performance was remarkably tolerant of a limited quantity of RS-fMRI data.

Keywords: convolutional neural network; deep learning; functional MRI; language; resting state network.