Convolutional Neural Network with Element-wise Filters to Extract Hierarchical Topological Features for Brain Networks

Xinying Xing, Junzhong Ji, Yao Yao

2018 · 37 citations · 14 references

Concepts

Abstract

Human brain network analysis based on machine learning has been paid much attention in the field of neuroimaging, where the application of convolutional neural network (CNN) is now becoming a new research hotspot. However, all present researches based on conventional CNN share weights on edges connected to the same node in a brain network, which ignores that each edge between any two nodes has a unique meaning and is not suitable for weight-sharing. In this paper, we propose a new convolutional neural network with element-wise filters (CNN-EW) for brain networks. More specifically, each element-wise filter gives a unique weight to each edge of brain network which may reflect the topological structure information more realistically. The experimental results on the autism brain imaging data exchange I (ABIDE I) dataset show that CNN-EW models can not only more accurately distinguish subject groups compared to some fashionable methods but also identify the abnormal brain regions associated with autism spectrum disorder (ASD).

References

14