[论文精读]DynBrainGNN: Towards Spatio-Temporal Interpretable Graph Neural Network Based on Dynamic Brain

论文全名：DynBrainGNN: Towards Spatio-Temporal Interpretable Graph Neural Network Based on Dynamic Brain Connectome for Psychiatric Diagnosis

论文网址：DynBrainGNN: Towards Spatio-Temporal Interpretable Graph Neural Network Based on Dynamic Brain Connectome for Psychiatric Diagnosis | 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

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. Proposed Model

2.3.1. Problem Definition

2.3.2. Overall Framework of DynBrainGNN

2.3.3. Construction of Dynamic Functional Graph

2.3.4. Graph Encoder

2.3.5. Spatio-Temporal Attention-Based READOUT Module

2.3.6. Dynamic Variational Autoencoders (DVAE)

2.4. Experiments

2.4.1. Dataset

2.4.2. Baselines

2.4.3. Experimental Settings

2.4.4. Evaluation on Classification Performance

2.5. Interpretation Analysis

2.5.1. Disease-Specific Brain Dynamic Network Connections

2.5.2. Temporal Properties

2.5.3. Conclusion

3. 知识补充

3.1. Dwell time

4. Reference List

---

1. 省流版

1.1. 心得

（1）完了写完了才发现没有心得，那我咋总结啊？

1.2. 论文总结图

2. 论文逐段精读

2.1. Abstract

        ①Again, FC can not present the dynamic character of fMRI

elucidate  v.阐明;说明;解释

2.2. Introduction

        ①⭐They challenge that the exist dynamic models lack interpretability in dwell time, fractional windows and number of transitions（意思你能解决？）

        ②They propose a Dynamic Brain Graph Neural Networks (DynBrainGNN), which based on dynamic brain connectom via dynamic variational autoencoders (DVAE) and spatio-temporal attention

        ③It is the first time that someone put forward such a "build in" dynamic FC?（啥玩意？真真第一次？）

2.3. Proposed Model

2.3.1. Problem Definition

        ①Graph set: , where  is the time series with  length of the -th subject and  is the number of subjects

        ②Through graphs, they extract and learn features 

        ③The real label set: 

2.3.2. Overall Framework of DynBrainGNN

        ①The schematic of DynBrainGNN:

which covers graph encoder, spatial attention module, temporal attention module and DVAE four modules;

where decoder recovers  and ;

然后，作者只是说橙色蓝色那俩框框是“为了保证解码器的可靠”，然后也没多说了

2.3.3. Construction of Dynamic Functional Graph

        ①Length of time series: 

        ②Length of slicing window: 

        ③Stride: 

        ④By dynamic cutting, they obtain  windowed dFC matrices（为啥？如果T=10，L=8，S=1，W不就等于2了吗，但是看上去是不是有1-8，2-9，3-10三个啊，不会要加一吗？）

        ⑤Each dFC calculated by Pearson correlation

        ⑥⭐They get the graph  where  is a adjacency matrix that all 1 transformed by the top 20% absolute correlation and  denotes the node features which constructed by the row or column of FC matrix

2.3.4. Graph Encoder

        ①Graph encoder: GCN

        ②Propagation rule of GCN:

where ,  denotes learnable parameters and  denotes Sigmoid

2.3.5. Spatio-Temporal Attention-Based READOUT Module

        ①They designed two attention based READOUT methods, Spatial Attention READOUT (SAR) and Temporal Attention READOUT (TAR)

        ②Based on prior , they define 

        ③In SAR, , where  denotes concatenation

        ④In TAR,  is constructed by the concatenation of several graph presentations at different times

        ⑤The specific operation of :

after Sigmoid, . "Then, attention masks are sampled from Bernoulli distributions, and the gumbelsoftmax reparameterization trick is applied to update "

        ⑥In SAR, 

        ⑦In TAR,  where  denotes Kronecker product

        ⑧Schematic of SAR and TAR:

2.3.6. Dynamic Variational Autoencoders (DVAE)

        ①Temporal transition: 

        ②The function of DVAE:

where ,  represents the encoder model（什么东西啊？就是GCN吗？）,  indicates the Frobenius norm,  denotes the prior distribution with isotropic Gaussian (assumed),  and  are both  scaling coefficients of the regularization term

        ③One more regularization term for compacting:

where  denotes the matrix-based Renyi’s -order mutual information and  denotes the scaling coefficient

        ④Accordingly, combining them all  to get a final loss function, where  is cross entropy loss

2.4. Experiments

2.4.1. Dataset

        ①ABIDE I: 289 ASD and 23 HC for no reason

        ②REST-meta-MDD: 397 MDD and 427 HC

        ③SRPBS: " This is a multi-disorder MRI dataset"（吓我一跳，总感觉是同时身患玉玉症多动症焦虑症自闭症老年痴呆的被试呢）, selecting 234 SCZ and 92 HC

2.4.2. Baselines

        ①Settings:

2.4.3. Experimental Settings

        ①Cross validation: 5 fold

        ②Decision of hyper-parameter: grid search

2.4.4. Evaluation on Classification Performance

        ①Comparison table:

2.5. Interpretation Analysis

2.5.1. Disease-Specific Brain Dynamic Network Connections

        ①The interpretations of dynamically dominant and fluctuant connections（？）are brought by  and 

        ②They define dominant subgraph  and fluctuant subgraph :

where  and  is the mean value of 

        ③The top 50 influential edges:

sensorimotor  adj. 感觉运动的（等于 sensomotor）

2.5.2. Temporal Properties

        ①“我们提供的时间属性的解释，以了解大脑的灵活性和适应性在精神疾病。具体而言，我们首先应用k-means聚类算法对有窗时空参与的图表示hGT进行聚类，以评估动态大脑模式(状态)。使用基于轮廓分数的聚类有效性分析来确定最佳聚类数量。然后，我们量化这些状态的时间属性的组差异，包括停留时间(即属于一个状态的连续窗口的持续时间)，分数窗口(即属于一个状态的总窗口的比例)和转换数量(即状态之间的转换数量)。使用带有错误发现率(FDR)校正的双样本t检验(图4)。我们的分析显示，ASD患者在II状态下有更高的分数窗口和平均停留时间，这与最近的一项神经影像学研究一致”（我失去了paraphrase能力）

        ②Temporal properties:

2.5.3. Conclusion

        They want to further try their model in other datasets

3. 知识补充

3.1. Dwell time

搜了一圈没搜到关于医学设备的，提供以下猜测

（1）最可能的，length of time series signals

（2）两个相邻点之间的时间？比如task fMRI两次task之间的时间间隔

（3）嘻嘻，事实证明上面俩都是错的，在2.5.2.作者说它是"the duration of consecutive windows belonging to one state"

4. Reference List

Zheng, K., Ma, B. & Chen, B. (2024) 'DynBrainGNN: Towards Spatio-Temporal Interpretable Graph Neural Network Based on Dynamic Brain Connectome for Psychiatric Diagnosis', Machine Learning in Medical Imaging, 14349.doi: DynBrainGNN: Towards Spatio-Temporal Interpretable Graph Neural Network Based on Dynamic Brain Connectome for Psychiatric Diagnosis | SpringerLink