[论文精读]Identification of Autism Spectrum Disorder With Functional Graph Discriminative Network

论文网址：Identification of Autism Spectrum Disorder With Functional Graph Discriminative Network - PMC (nih.gov)

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

目录

1. 省流版

1.1. 心得

1.2. 论文总结图

2. 论文逐段精读

2.1. Abstract

2.2. Introduction

2.3. Materials and Methods

2.3.1. Graph Construction

2.3.2. Proposed Graph Model

2.4. Results

2.4.1. ABIDE Data Preprocessing

2.4.2. Experiments

2.5. Discussion

2.6. Conclusion

3. Reference

---

1. 省流版

1.1. 心得

（1）摘要没说方法，只说了应用上

（2）intro也没咋写方法，说贡献是提出了一个模型hhhh，真会说话

（3）公式写得太冗杂了，没必要重复很多次，下标也太多了

（4）我怎么没理解它FC是怎么来的？似乎没提到时间序列和皮尔逊什么的，通用吗？随意吗？是我没看到吗？

（5）分别用5 fold和10 fold分别凑一张对比表.......我 啊这

（6）医学部分也没好好说，感觉能二区大概是实验做的多吧（虽然也没有多大用）

1.2. 论文总结图

2. 论文逐段精读

2.1. Abstract

        ①Facing the heterogeneity of ASD, traditional functional connectivity (FC) methods can not accurately classify ASD and HC

        ②They proposed a functional graph discriminative network (FGDN)

2.2. Introduction

        ①Introducing ASD and the meaning of pre-diagnosis for ASD patients

        ②Noting how fMRI applies to ASD diagnosis

        ③Listing some machine learning methods on ASD diagnosis（是比较早期的，所以倒不至于和related work重复）

        ④Further introducing and explain the graph representation methods

        ⑤Their contributions: a) proposed FGDN, b) high classification performance, c) researched biomarkers

2.3. Materials and Methods

2.3.1. Graph Construction

        ①Graph construction: 

where  and  denote node set and edge set in graph ,  denotes the adjacency matrix（可能需要特别注意一下，因为其他的一般都用）. And the adjacency matrix constructed by Gaussian distance function and k-nearest neighbor:

        ②FC matrix construction:

where  denotes the class (0 means ASD and HC is 1)

        ③Two templates:

with

2.3.2. Proposed Graph Model

        ①用了谱图卷积和一阶近似...什么的。但是感觉也是用的别人的方法，又再次介绍了一下。我就不赘述了

        ②The overall framework:

        ③Atlas: AAL 116, HO 118 and MODL 128

        ④Graph kernel: 64

        ⑤Order of Chebyshev: 3

        ⑥Activation function: Sigmoid

        ⑦Loss function: cross entropy

2.4. Results

2.4.1. ABIDE Data Preprocessing

        ①Sample: 866 with 402 ASD and 464 HC

        ②FC: Ledoit–Wolf regularized shrinkage estimator（对什么数据使用啊？？？？）

        ③ in NN: 20

2.4.2. Experiments

        ①Validation: 5 or 10-fold cross validation

        ②Optimizer: Adam

        ③Learning rate: 0.0001

        ④Weight decay: 0.0005

        ⑤Batch size: 16

        ⑥Dropout rate: 0.1

        ⑦5 fold comparison:

        ⑧10-fold comparison:

2.5. Discussion

        ①Loss and training curve（感觉也没什么用）:

        ②Training ratio（一般来说增加训练样本都更好啦...）:

        ③Leave-one-site-out performance:

        ④5 most discriminative brain regions by t-test:

2.6. Conclusion

        ......

3. Reference

Li, J. et al. (2021) 'Identification of Autism Spectrum Disorder With Functional Graph Discriminative Network', Frontiers in Neuroscience. doi: 10.3389/fnins.2021.729937