[BIBM 2021]A Graph Attention Neural Network for Diagnosing ASD with fMRI Data

最新推荐文章于 2025-04-03 12:40:48 发布

夏莉莉iy

最新推荐文章于 2025-04-03 12:40:48 发布

阅读量779

点赞数 15

分类专栏：论文精读文章标签：深度学习人工智能机器学习算法神经网络图论 python

本文链接：https://blog.youkuaiyun.com/Sherlily/article/details/144806833

版权

论文精读专栏收录该内容

152 篇文章

订阅专栏

论文网址：A Graph Attention Neural Network for Diagnosing ASD with fMRI Data | IEEE Conference Publication | IEEE Xplore

英文是纯手打的！论文原文的summarizing and paraphrasing。可能会出现难以避免的拼写错误和语法错误，若有发现欢迎评论指正！文章偏向于笔记，谨慎食用

2.3. Materials and Methods

2.3.1. Preliminaries on graph attention network for classification

2.3.2. Functional connectivity derived graph and feature construction

2.3.3. Connectivity based GAT model for Autism classification

2.4. Experiments and Results

2.4.1. Dataset preparation

2.4.2. Experiment setup

2.4.3. Discussions

2.5. Conclusion

3. Reference

1. 心得

（1）考古学家会梦到新idea吗

（2）有一种在刷全收集的感觉

（3）难以想象我如果21年读研会有多么开心

2. 论文逐段精读

2.1. Abstract

①They proposed a connectivity based graph attention network for autism diagnosis using functional connectivity (FC)（摘要也没说现在有什么挑战，可能因为太早了，还在拓展到临床应用）

2.2. Introduction

①很传统的intro，从ASD到fmri到FC和应用再到他们自己的GNN

2.3. Materials and Methods

2.3.1. Preliminaries on graph attention network for classification

①Representation of directed graph: $G=(V,E,W)$ , where $V=1,\cdots,N$ denotes node set, $E\in V\times V$ is edge set, and $W{\in}[0,1]^{|V|\times|V|}$ denotes adjacency matrix（谁用 $W$ 表示邻接矩阵啊）

②Node feature: $h_i^0\in R^d$

③For input nodes $\mathbf{h}=\left\{\vec{h}_1,\vec{h}_2,\ldots,\vec{h}_N\right\},\vec{h}_i\in R^F$ , GAT calculates attention coefficients:

$e_{ij}=a \begin{pmatrix} \mathbf{W}\vec{h}_i,\mathbf{W}\vec{h}_j \end{pmatrix}$

where $\mathbf{W}$ is learnable weight matrix

④Apply Softmax on attention score:

$\alpha_{ij}=\mathrm{softmax}_j(e_{ij})=\frac{\exp(e_{ij})}{\sum_{k\in\mathcal{N}}\exp(e_{ik})}$

⑤Updating function of single head GAT:

$h_i^{\prime}=\sigma\left(\sum_{j\in\mathcal{N}_i}\alpha_{ij}\mathbf{W}\vec{h}_j\right)$

where $\sigma$ denotes activation function

⑥Multi head GAT:

$\begin{aligned} & h_{i}^{\prime}=\|_{k=1}^K\sigma\left(\sum_{j\in N_i}a_{ij}^k\mathbf{W}^k\mathbf{h}_j\right) \\ & h_{i}^{\prime}=\sigma\left(\frac{1}{K}\sum_{k=1}^K\sum_{j\in N_i}a_{ij}^k\mathbf{W}^k\mathbf{h}_j\right) \end{aligned}$

where $\parallel$ denotes concatenation operation

2.3.2. Functional connectivity derived graph and feature construction

①Diagnose ASD by FC:

②Pearson correlation (PC) function:

$r_{xy}=\frac{\sum_{i=1}^L\left(x_i-\bar{x}\right)\left(y_i-\bar{y}\right)}{\sqrt{\sum_{i=1}^L\left(x_i-\bar{x}\right)^2}\sqrt{\sum_{b=1}^L\left(y_b-\bar{y}\right)^2}}$

③Parcellation: Powell 264

④Rewrite undirected graph as: $G=(V,E,A)$

⑤7 node feature: a) 3 graph related feature: centralities-graph degree, graph betweenness, graph eccentricity; b) 4 time related: mean, variance, skewness and kurtosis

skewness n.偏斜 kurtosis n.峰度；（曲线的）峰态；峭度；突出度（分布曲线中的高峰程度）；（统）尖削度

2.3.3. Connectivity based GAT model for Autism classification

①PC is limited in high-order relationship capturing

2.4. Experiments and Results

2.4.1. Dataset preparation

①Dataset: ABIDE I

②Preprocessing: standard Open Tools, ANFI and FSL

2.4.2. Experiment setup

①Network parameters:

each GAT layer follows a ReLU layer, classifier is MLP with SoftMax, batch size is 8, dropout rate is 0.5

②Sample number: 871

③Data split: 75%:25%

④Cross validation: 5 fold

⑤Performance table:

2.4.3. Discussions

①Just a test

2.5. Conclusion

3. Reference

Yin, W. et al. (2021) A Graph Attention Neural Network for Diagnosing ASD with fMRI Data, IEEE International Conference on Bioinformatics and Biomedicine (BIBM), Houston, TX, USA.