Deep Neural Networks for Assessing Functional Connectivity: an fNIRS Study

Abstract

Studies on interactions between brain regions estimate functional connectivity which are usually based on the basis of temporal presence. Functional connectivity derived from resting-state has been attracted by several recent studies as it provides valuable insight into the intrinsic networks of the human brain. Functional near-infrared spectroscopy (fNIRS) has gained attention in resting-state functional connectivity (RSFC) patterns detection because of its advantages compared to other neuroimaging modalities. Several progressive methodologies in detecting RSFC patterns in fNIRS, such as seed-based correlation analysis, and independent component analysis (ICA), were adopted in previous studies. Despite the fact that it is not known which methodology is the most suitable in detecting RSFC patterns, seed-based correlation analysis and ICA-based analysis which are the most widely used methodologies in RSFC studies, have intrinsic disadvantages. Therefore, in this study a method based on artificial neural network(ANN) was introduced to meet the possibilities of overcoming the conventional methods challenges.

The RSFC patterns of the sensorimotor system derived from ANN were consistent with the previous findings. Moreover, the results of ANN illustrated the superior performance in the terms of specificity and sensitivity compared to both conventional approaches. The main contribution of the present thesis is to emphasize that ANN can be used as an appropriate method to estimate the temporal relation among brain networks during resting-state. ⓒ 2017 DGIST