Elsevier

Pattern Recognition

Volume 109, January 2021, 107570
Pattern Recognition

Convolutional kernels with an element-wise weighting mechanism for identifying abnormal brain connectivity patterns

https://doi.org/10.1016/j.patcog.2020.107570Get rights and content

Highlights

  • New CKEW can extract hierarchical topological features for brain networks.

  • Our feature analysis method can find key features of brain networks as biomarkers.

  • CKEW can obtain the highest classification accuracies among all testing methods.

Abstract

Deep learning based human brain network classification has gained increasing attention in recent years. However, current methods remain limited in exploring the topological structure information of a brain network. In this paper, we propose a kind of new convolutional kernels with an element-wise weighting mechanism (CKEW) to extract hierarchical topological features of brain networks, in which each weight is assigned to an element with a unique neuroscientific meaning. In addition, a novel classification framework based on CKEW is presented to diagnose brain diseases and explore the most important original features by a tracing feature analysis method efficiently. Experimental results on two autism spectrum disorder (ASD) datasets and an attention deficit hyperactivity disorder (ADHD) dataset with functional magnetic resonance imaging (fMRI) data demonstrate that our method can more accurately distinguish subject groups compared to several state-of-the-art methods in cerebral disease classification, and abnormal connectivity patterns and brain regions identified are more likely to become biomarkers associated with a cerebral disease.

Introduction

In the remarkable field of brain science, more and more researches on brain cognition and diagnosis of cranial nerve diseases are beginning to explore the origins of human intelligence and pathogenesis of various cerebral diseases based on a wide range of brain imaging data. The investigation of brain networks provides an opportunity to explore the relationships between abnormal connectivity patterns of the human brain and cranial nerve diseases, which helps understand the pathological basis of neurological disorder and is essential for the early diagnosis of neuropsychiatric disorders.

Brain networks can be classified into two main categories: structural connectivity networks and functional connectivity networks. This paper focuses on functional connectivity networks based on resting-state functional Magnetic Resonance Imaging (rs-fMRI) due to its ability to reveal the patterns of brain functional activity. Functional connectivity networks can also refer to the functional connectome, and the two concepts are used alternatively in the following. A functional connectivity network consists of nodes and edges, where a node is corresponding to a neuron, or a brain region of interest (ROI) on a macroscopic scale, and each edge represents the correlation measure (e.g. Pearson correlation coefficient) between a pair of nodes. It has been demonstrated that functional connectivity is an effective biomarker for the diagnosis of brain diseases [1], [2] and has been widely applied, where the functional connectivity features were extracted and classified using different types of methods.

Traditional approaches applied in functional connectivity networks, such as support vector machine (SVM) and linear discriminant analysis (LDA), highly rely on the handcrafted feature based methods. These methods only extract low-level features by a simple linear combination and nonlinear transformation of raw feature (functional connectivity), and they are limited in feature representation and classification.

Recently, an increasing number of researches have introduced deep neural networks (DNN) to improve the classification of brain networks. The DNNs can extract high-level features by means of some deep structures, from which the high-level representation of raw feature is able to be further explored and learned. Fully connected neural network (FCNN) and convolutional neural network (CNN) are two typical categories in the DNN based brain network classification methods, which have shown natural superiority in high-level feature learning. FCNN methods usually require a large number of parameters learning which is time-consuming, and are prone to overfitting in dealing with the small sample of brain network data with high dimensions. CNN methods employ local connections and weight-sharing mechanisms to extract high-level features from grid-like data, and attract more attention than FCNN methods. Because the topological locality of each edge in a brain network data is different from the spatial locality of each pixel in an image data, the current CNN methods including BrainNetCNN [3] and connectome-CNN (CCNN) [4] designed new feature extraction structures according to the topological locality characteristics of brain networks. Recently, Santana et al. had conducted systematical experiments and demonstrated that these deep learning models had a better performance compared to other less costly models such as SVM and RFC [5]. However, these methods still employed the typical convolutional kernels to perform convolution processes by row or column. In fact, each edge in a brain network uniquely denotes a particular connectivity between a pair of brain regions, and has a specific neuroscientific meaning and role [6], which makes it unreasonable to share weights on the edges connected to the same node by typical convolutional kernels.

To efficiently extract deep features of the functional connectome and take advantage of its topological structure information, we propose a kind of new convolutional kernels with an element-wise weighting mechanism (named CKEW) for extracting the hierarchical topological features of brain networks and then apply it to the cerebral disease classification. The CKEW gives a unique weight to each element (edge) in a brain connectome, rather than shares weights on all edges connected to the same node as done in related works [3], [4]. Thus it can more effectively extract the topological structure information of the brain network, which will improve the performance of brain network classification significantly. The main contributions of this paper are as follows:

  • (1)

    According to the specific topological locality of the functional connectome, we design a kind of convolutional kernels with an element-wise weighting mechanism, where each weight is assigned to an element with a unique neuroscientific meaning, and then we propose a new classification framework based on CKEW to extract hierarchical topological features for the brain network.

  • (2)

    Based on the proposed framework, we propose a tracing feature analysis method on the basis of edges and nodes in the brain network by accumulating the importance of higher-level features from top to down, which is apt to find the significant relevance between key original features of the brain network and their classification results, and identify biomarkers associated with a brain disease.

  • (3)

    Systematic experiments have been conducted to verify the proposed CKEW and the new framework on a series of simulated datasets and three real datasets. The experimental results demonstrate the effectiveness of CKEW and the proposed framework on all testing cases. And our method is able to identify important function connectivities and brain regions significantly.

This paper is organized as follows. Section 2 summarizes the related works briefly. Section 3 introduces the proposed CKEW model and classification framework. Section 4 analyzes the experimental results. Section 5 concludes this paper and outlines the future research work.

Section snippets

Brain networks

Recently, many pattern recognition approaches have been successfully applied into brain cognition and diagnosis of cranial nerve diseases. Remarkably, since the Human Connectome Project (HCP) launched in 2009, more and more researches indicated that many brain diseases are usually accompanied by the interruption or abnormal integration of connections between brain regions [7]. Zhou et al. proposed two general discriminative learning frameworks for Gaussian Bayesian network (BN) to improve the

Main idea

A brain network is typically represented as a graph G=(V,E), where V is a set of nodes represented brain regions and E is an adjacency matrix where each element (edge) EijE represents the connection strength between each pair of brain regions Vi and Vj. Though the brain network, represented as an adjacency matrix, is the grid-like data, there are two essential differences between the brain network data and the image data. First, the topological locality of each element (edge) in a brain

Experiments

In this section, we use simulated datasets and real fMRI datasets of ASD to perform our empirical study. The experimental platform is a PC with Inter Core i7-8700k, 3.70 GHz CPU, NVIDIA GeForce GTX 1080Ti GPU, and Ubuntu 16.04.

Conclusions

In this paper, we present a kind of novel convolutional kernels with an element-wise weighting mechanism (called CKEW), where each weight is assigned to an edge with a unique neuroscientific meaning in the brain network, to extract hierarchical topological features of functional connectivity networks. Experimental results validate the superiority of the proposed CKEW-based classification framework in accurately capturing the inherent features of functional connectivity networks. Additionally,

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgment

This work is partly supported by NSFC Research Program (61672065 and 61906007). Part of data collection and sharing for this project was funded by the R-fMRI Maps Project.

JunZhong Ji is a professor, Ph.D. supervisor at Faculty of Information Technology, Beijing University of Technology, Youth Skeleton teacher in Beijing and senior membership of China Computer Federation. His research interests include Data Mining, Machine learning, Swarm Intelligence and Bioinformatics. He has published over 60 papers in these areas.

References (45)

  • F. Pereira et al.

    Machine learning classifiers and fMRI: a tutorial overview

    Neuroimage

    (2009)
  • V. Menon

    Large-scale brain networks and psychopathology: a unifying triple network model

    Trends Cogn. Sci.

    (2011)
  • X. Duan et al.

    Resting-state functional under-connectivity within and between large-scale cortical networks across three low-frequency bands in adolescents with autism

    Prog. Neuro-Psychopharmacol. Biol. Psychiatry

    (2017)
  • Y. Agam et al.

    Reduced cognitive control of response inhibition by the anterior cingulate cortex in autism spectrum disorders

    Neuroimage

    (2010)
  • P. Shih et al.

    Atypical network connectivity for imitation in autism spectrum disorder

    Neuropsychologia

    (2010)
  • N. de Lacy et al.

    Disruption to control network function correlates with altered dynamic connectivity in the wider autism spectrum

    NeuroImage Clin.

    (2017)
  • M. Solomon et al.

    The neural substrates of cognitive control deficits in autism spectrum disorders

    Neuropsychologia

    (2009)
  • X. Liang

    Human connectome: structural and functional brain networks

    Chin. Sci. Bull.

    (2010)
  • T. Itahashi et al.

    Altered network topologies and hub organization in adults with autism: a resting-state fMRI study

    PLoS ONE

    (2014)
  • R.J. Meszlényi et al.

    Resting state fMRI functional connectivity-based classification using a convolutional neural network architecture

    Front. Neuroinform.

    (2017)
  • A.N. Santana et al.

    Using deep learning and resting-state fMRI to classify chronic pain conditions

    Front. Neurosci.

    (2019)
  • D. Zhang et al.

    Machine-learning-based brain network analysis: method and application

    J. Data Acquis. Process.

    (2015)
  • Cited by (0)

    JunZhong Ji is a professor, Ph.D. supervisor at Faculty of Information Technology, Beijing University of Technology, Youth Skeleton teacher in Beijing and senior membership of China Computer Federation. His research interests include Data Mining, Machine learning, Swarm Intelligence and Bioinformatics. He has published over 60 papers in these areas.

    Xinying Xing received her master’s degree at Faculty of Information Technology, Beijing University of Technology. Her main research interests focus on deep learning and bioinformatics.

    Yao Yao received the bachelor’s degree from Beijing University of Technology and is currently working toward the PhD degree in computer science at the Beijing University of Technology. His main interest include computation intelligence, deep learning and brain science.

    Junwei Li is studying for a doctor’s degree at Faculty of Information Technology, Beijing University of Technology. Her research interests focus on deep learning, ensemble learning and bioinformatics.

    Xiaodan Zhang is a postdoctoral fellow in Faculty of Information Technology, Beijing University of Technology. Her current research interests include computer vision, natural language processing, image captioning, and deep learning.

    View full text