FCNet: A Convolutional Neural Network for Calculating Functional Connectivity from Functional MRI

Abstract

This is the accepted version of the paper.This version of the publication may differ from the final published version.Permanent repository link: http://openaccess.city.ac.uk/18045/ Link to published version: http://dx.Abstract. Investigation of functional brain connectivity patterns using functional MRI has received significant interest in the neuroimaging domain. Brain functional connectivity alterations have widely been exploited for diagnosis and prediction of various brain disorders. Over the last several years, the research community has made tremendous advancements in constructing brain functional connectivity from timeseries functional MRI signals using computational methods. However, even modern machine learning techniques rely on conventional correlation and distance measures as a basic step towards the calculation of the functional connectivity. Such measures might not be able to capture the latent characteristics of raw time-series signals. To overcome this shortcoming, we propose a novel convolutional neural network based model, FCNet, that extracts functional connectivity directly from raw fMRI time-series signals. The FCNet consists of a convolutional neural network that extracts features from time-series signals and a fully connected network that computes the similarity between the extracted features in a Siamese architecture. The functional connectivity computed using FCNet is combined with phenotypic information and used to classify individuals as healthy controls or neurological disorder subjects. Experimental results on the publicly available ADHD-200 dataset demonstrate that this innovative framework can improve classification accuracy, which indicates that the features learnt from FCNet have superior discriminative power.

Deep fMRI: AN end-to-end deep network for classification of fMRI data

“…The network is trained end-to-end. For initialization of the feature extractor and similarity measure networks, we use weights from a pretrained FCNet in our work [7], and these weights are updated through fine-tuning. The end-to-end model is trained with the following loss:…”

“…The proposed work is motivated by a recently published method called FCNet [7]. FCNet is used to extract functional connectivity from fMRI time-series signals, but it suffers from the following drawbacks: i) it is not an end-to-end model, and ii) it relies on classical machine learning methods like feature selection using elastic net and a support vector machine for classification.…”

“…Instead of initializing weights of the feature extractor network and similarity measure network randomly, we use weights of a pre-trained FCNet [7].…”

“…Generally speaking, an end-to-end trainable network refers to a single learning system where the predicted label of a machine learning process is predicted directly from the raw input, with all weights learned through back-propagation. Recently, a deep learning method named FCNet [7] has been proposed for classification of ADHD from fMRI data. The method uses a convolutional neural network (CNN) to predict functional connectivity of brain regions.…”

“…1. The proposed model is inspired by the FCNet [7] and uses a pretrained FCNet for some of the initial layers to extract features from the raw fMRI time-series signals. Unlike FCNet, these layers are fine-tuned during training to learn the features of individual brain regions.…”