Language task-based fMRI analysis using machine learning and deep learning

Abstract

Introduction: Task-based language fMRI is a non-invasive method of identifying brain regions subserving language that is used to plan neurosurgical resections which potentially encroach on eloquent regions. The use of unstructured fMRI paradigms, such as naturalistic fMRI, to map language is of increasing interest. Their analysis necessitates the use of alternative methods such as machine learning (ML) and deep learning (DL) because task regressors may be difficult to define in these paradigms.

Methods: Using task-based language fMRI as a starting point, this study investigates the use of different categories of ML and DL algorithms to identify brain regions subserving language. Data comprising of seven task-based language fMRI paradigms were collected from 26 individuals, and ML and DL models were trained to classify voxel-wise fMRI time series.

Results: The general machine learning and the interval-based methods were the most promising in identifying language areas using fMRI time series classification. The geneal machine learning method achieved a mean whole-brain Area Under the Receiver Operating Characteristic Curve (AUC) of $0.97\pm 0.03$ , mean Dice coefficient of $0.6\pm 0.34$ and mean Euclidean distance of $2.7\pm 2.4$ mm between activation peaks across the evaluated regions of interest. The interval-based method achieved a mean whole-brain AUC of $0.96\pm 0.03$ , mean Dice coefficient of $0.61\pm 0.33$ and mean Euclidean distance of $3.3\pm 2.7$ mm between activation peaks across the evaluated regions of interest.

Discussion: This study demonstrates the utility of different ML and DL methods in classifying task-based language fMRI time series. A potential application of these methods is the identification of language activation from unstructured paradigms.

Keywords: brain activation; deep learning; language; machine learning; task-based fMRI; time series.

Figure 1

Illustrated are the steps to obtain task-based language activation maps. Task paradigm stimulus function was convolved with the hemodynamic response function to get task regressors. Raw fMRI images were preprocessed, and together with task and nuisance regressors activation maps are produced using Genearl Linear modelling.

Figure 2

Illustrated are the steps to extract voxel time series data and corresponding binary labels. Task-based language activation maps were used to define the 0 and 1 labels.

Figure 3

This figure shows the overlap between SPM activation maps vs. evaluated ML/DL activation maps of two test participants (A healthy participant, LHS under each algorithm title and an epilepsy patient, RHS under each algorithm title) for a single single language task - Sentence Completion (SC). Black - Overlap, yellow - SPM activation, Red - ML/DL activation.

Figure 4

Mean whole-brain AUC values across test participants of different ML/DL categories, by language paradigm (as a scatter plot with violin plots showing the distribution of AUC values for the best and worse performing ML/DL methods for each language paradigm. Blue - best, Orange - worst).

Figure 5

Mean Dice coefficient across test participants of different ML/DL categories, by language regions of interest (as a scatter plot with violin plots denoting the distribution of Dice coefficient values for the best and worse performing ML/DL methods for each language paradigm. Blue - best, Orange - worst).

Figure 6

Mean Euclidean distance between activation peaks across test participants of different ML/DL categories and SPM, by language regions of interest (as a scatter plot with violin plots denoting the distribution of Euclidean distances for the best and worse performing ML/DL methods for each language paradigm. Blue - best, Orange - worst).