[论文精读]Intelligence Quotient Scores Prediction in rs-fMRI via Graph Convolutional Regression Network

最新推荐文章于 2024-06-26 16:34:42 发布

夏莉莉iy

最新推荐文章于 2024-06-26 16:34:42 发布

阅读量859

点赞数 24

分类专栏：论文精读文章标签：人工智能机器学习深度学习计算机视觉学习笔记图论

本文链接：https://blog.csdn.net/sherlily/article/details/136081803

版权

论文精读专栏收录该内容

54 篇文章 8 订阅

订阅专栏

论文网址：Intelligence Quotient Scores Prediction in rs-fMRI via Graph Convolutional Regression Network | SpringerLink

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！

2.3.2. Proposed Network

2.4. Experimental Results

2.4.1. Experimental Settings

2.4.2. Performance Evaluation

2.4.3. Biomarker Detection

2.5. Conclusion

3. 知识补充

3.1. Poly learning rate

4. Reference List

1. 省流版

1.1. 心得

（1）用自闭症数据集来预测智商？？好吧（我真的想说这后面也没有圆回来！！！我还以为有什么惊为天人的发现呢！）

（2）从前面看是好简单的一个网络啊！后面看也没有很难

（3）这个随机选样本，hhhhh

（4）你有什么拓扑结构？？？我怎么没get到。提取的特征就是高级拓扑特征？

（5）宝宝你是一个...

1.2. 论文总结图

2. 论文逐段精读

2.1. Abstract

①Due to the complexity of brain connections and the change of topological structure in brain, IQ predictions always get inaccurate results

②They put forward a model including attention branch and global branch

2.2. Introduction

①There is no relevant topic of IQ prediction in ASD datasets

②Machine learning separate feature selection and regression. However, it is hard to obrain accurate advanced topological features from brain networks

③They proposed a Graph Convolutional Regression Network (GCR-Net) for IQ prediction

engrain v.使根深蒂固，深透，确立 adj.根深蒂固的

2.3. Methodology

2.3.1. Pipeline

①The schematic of GCR-Net

②Atlas: Automated Anatomical Labeling (AAL) with 116 ROIs

③Functional connectivity matrix : calculated by Pearson correlation $\mathbf{A}\in\mathbb{R}^{n\times n}$

④The brain graph $G=\{\mathbf{V},\mathbf{A}\}$ , where $\mathbf{V}=\{v_1,\ldots,v_n\}$ are ROIs

2.3.2. Proposed Network

①Their overall framework:

where $u$ denotes the node features

②The ConvNet includes a graph convolution (GC) layer and a dropout layer

（1）Attention branch

①The operation of GC:

$\begin{aligned} \mathbf{X}^{(l)}& =\text{Re}\mathrm{LU}(\mathrm{GC}(\mathbf{X},\mathbf{A}) \\ &=\operatorname{ReLU}\left(\mathbf{D}^{-\frac12}\mathbf{A}\mathbf{D}^{-\frac12}\mathbf{X}^{(l-1)}\mathbf{W}^{(i)}\right) \end{aligned}$

where $\mathbf{D}=\operatorname{diag}\left(\sum_j\mathbf{A}(1,j),\sum_j\mathbf{A}(2,j),\ldots,\sum_j\mathbf{A}(n,j)\right)$ represents the degree matrix of $\mathbf{A}$ and they set $\mathbf{A}^{(0)}=\textbf{I}$ （等等，为什么我这里怎么也理解不了，这A是个啥玩意儿啊？）;

$\mathbf{X}^{(l)}=\left[x_1^{(l)},x_2^{(l)},\ldots,x_n^{(l)}\right]^\top$ denotes node features

②In the self attention module, the attention score $\mathbf{S}\in\mathbb{R}^{r\times n}$ can be calculated by:

$S=\text{softmax}\left(\mathbf{W}_{s2}\tanh(\mathbf{W}_{s1}\mathbf{H}^\top)\right)$

③The pooling module contains matrix multiplication $e=\mathbf{SH}$ and maximum pooling. After pooling, it generates prediction $\mathbf{\hat{X}}\in\mathbb{R}^{1\times1}$

（2）Global branch

①The GC operation:

$\mathbf{F}=\mathrm{GC}(\mathbf{E},\mathbf{A}),\mathbf{F}\in\mathbb{R}^{n\times1}$

②Transposing $\mathbf{F}$ and regrding it as a global (personal) feature

③MLP $\mathbf{X}=\mathbf{W}_{s3}\mathbf{F}^\top,\mathbf{X}\in\mathbb{R}^{1\times1}$

（3）Finally, fusing $\mathbf{\hat{X}}$ and $\mathbf{X}$ with element-wise mean operation then giving the prediction

2.4. Experimental Results

2.4.1. Experimental Settings

（1）Settings

①Optimizer: Adam

②Learning rate: 0.001

③Weight decay: 0.0005

④⭐Poly learning rate: 0.9 power

⑤Maximum epoch: 100

⑥Loss function: mean square error (MSE)

⑦Hyper-parameters: $u=64,v=2,r=4$

（2）Dataset

①Autism Brain Imaging Data Exchange (ABIDE) adopted

②Preprocessing pipelines: Connectome Computation System (CCS), Configurable Pipeline for the Analysis of Connectomes (CPAC), Data Processing Assistant for Resting-State fMRI (DPARSF), Neuroimaging Analysis Kit (NIAK)

③⭐Samples: randomly choosing 226 NT and 202 ASD

（3）Evaluation Metrics

①Cross validation: 3 fold, 2 for training and 1 for test

②Runs: mean of 3

③mean absolute error (MAE): $\mathrm{MAE}=\frac1m\sum_{i=1}^m|y_i-\hat{y}_i|$

④root mean squared error (RMSE): $\mathrm{RMSE}=\sqrt{\frac1m\sum_{i=1}^m\left(y_i-\hat{y}_i\right)^2}$

where $\hat{y}$ is the predicted answer and $y$ is the true label, $m$ denotes the number of patients

2.4.2. Performance Evaluation

①Regression comparison in ASD:

②Regression comparison in NT:

2.4.3. Biomarker Detection

①5 most influential ROIs for IQ on ASD:

②5 most influential ROIs for IQ on NT:

③Analysing the similarity between predicted associated brain regions and medical research

episodic adj.不定期的;情节性的;偶尔发生的;有许多片段的;由松散片段组成的

2.5. Conclusion

No need for that.

3. 知识补充

3.1. Poly learning rate

参考学习1：Pytorch几种常用的学习率调整方式_pytorch poly-CSDN博客

参考学习2：pytorch动态调整学习率之Poly策略_poly学习率-CSDN博客

4. Reference List

Zhang H. et al. (2022) 'Intelligence Quotient Scores Prediction in rs-fMRI via Graph Convolutional Regression Network', CAAI International Conference on Artificial Intelligence, pp. 477-488. doi: Intelligence Quotient Scores Prediction in rs-fMRI via Graph Convolutional Regression Network | SpringerLink

夏莉莉iy

关注

24
点赞
踩
21

收藏

觉得还不错? 一键收藏
0
评论
[论文精读]Intelligence Quotient Scores Prediction in rs-fMRI via Graph Convolutional Regression Network

计算机-人工智能-脑科学与类脑智能
复制链接

扫一扫