[论文精读]Multiview hyperedge-aware hypergraph embedding learning for multisite, multiatlas fMRI based

夏莉莉iy

已于 2024-05-08 09:19:57 修改

阅读量991

点赞数 30

分类专栏：论文精读文章标签： embedding 人工智能笔记机器学习学习 python 深度学习

于 2024-05-03 15:52:52 首次发布

本文链接：https://blog.csdn.net/Sherlily/article/details/138402465

版权

论文精读专栏收录该内容

52 篇文章 7 订阅

订阅专栏

论文全名：Multiview hyperedge-aware hypergraph embedding learning for multisite, multiatlas fMRI based functional connectivity network analysis

论文原文：Multiview hyperedge-aware hypergraph embedding learning for multisite, multiatlas fMRI based functional connectivity network analysis - ScienceDirect

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

2.3.1. Hypergraph and HGCN with FCN

2.3.2. The proposed CcSi-MHAHGEL

2.4. Experimental results

2.4.1. Dataset and data preprocessing

2.4.2. Experimental settings

2.4.3. Methods for comparison

2.4.4. Classification performance

2.4.5. Ablation study

2.4.6. Construction of the shared incidence matrix

2.4.7. Most discriminative ROIs discovered by CcSi-MHAHGEL

2.5. Discussions

2.5.1. Visualization of FCN embeddings

2.5.2. Influence of key hyperparameters

2.5.3. Generalization ability of the class-consistency and site-independence modules

2.5.4. Results of CcSi-MHAHGEL with three-atlas-based FCNs

2.5.5. Results of CcSi-MHAHGEL on the whole ABIDE

2.5.6. Influence of hyperedge construction schemes

2.5.7. Limitations and future work

2.6. Conclusion

3. Reference List

1. 省流版

1.1. 心得

（1）既然作者想做的事情是跨站点那么只跨4/21个ABIDE站点是不是没有那么有说服力啊（消融实验测了很多的，19个站点，精度差不多）

（2）超图没什么选择，其实也就是多聚合了一下

（3）做了好多实验

（4）名字取得好长捏，有点让人记不住。最近KAN爆火，感觉凡事还是得从简

1.2. 论文总结图

2. 论文逐段精读

2.1. Abstract

①They proposed Class-consistency and Site-independence Multiview Hyperedge-Aware HyperGraph Embedding Learning (CcSi-MHAHGEL) framework

②Existing problems: insufficient researches in multi-site and multi-atlas

③Graph construction: one hypergraph per subject

2.2. Introduction

①Shortcomings of GCNs: simple structure, inadequate connections between subjects, different brain graph distributions between different sites

②Atlas: AAL and HO

③Samples: 355 with 167 ASD and 188 HC from the bigest 4 sites from ABIDE

④怎么又是Grad-CAM

⑤Their contributions: proposed hypergraph based on multi-atlases, class-consistency and site-independence modules

⑥ $\left | \cdot \right |$ denotes the cardinality of a set（就是一个集合里面有几个数）

atrophy n.萎缩 vi.萎缩；衰退 cardinality n.基数；集的势(mathematics) the number of elements in a set or group (considered as a property of that grouping)

2.3. Methods

2.3.1. Hypergraph and HGCN with FCN

①Representation of hypergraph:

②Hypergtaph can be represented as $G=\left ( V, E, W \right )$ , where $V=\left \{ v_i \right \}^{\left | V \right |}_{i=1}$ denotes vertex set (ROI), $E=\left \{ e_i \right \}^{\left | E \right |}_{i=1}$ denotes hyperedge set. Each hyperedge $e_i$ has its own weight $w(e_i)$ , and $W=diag({w(e_i)}^{\left | E \right |}_{i=1}) \in \mathbb{R}^{\left | E \right | \times \left | E \right |}$ denotes diagonal matrix of the hyperedge weights.

③All the hyperedges satisfy $\bigcup ^{\left | E \right |}_{i=1} e_i = V$

④Incidence matrix of $G$ is $H \in \mathbb{R}^{\left | V \right | \times \left | E \right |}$ with 0 or 1

⑤Degree of $v_i$ can be calculated by $d(v_i)=\sum _{e_j \in E} w(e_j) h (v_i,e_j)$ （有边就将边权重*1加起来）, and its diagonal degree matrices $D_v=diag\left ( {d\left ( v_i \right )}^{\left | V \right |} _{i=1}\right ) \in \mathbb{R}^{\left | V \right | \times \left | V \right |}$

⑥Degree of $e_j$ can be $d(e_i)=\sum _{v_i \in V} h (v_i,e_j)$ and its diagonal degree matrices $D_e=diag\left ( {d\left ( e_j \right )}^{\left | E \right |} _{j=1}\right ) \in \mathbb{R}^{\left | E \right | \times \left | E \right |}$

⑦Hypergraph Laplacian matrix can be expressed as $\Delta \in \mathbb{R}^{\left | V \right | \times \left | V \right |}$ , where $\Delta=I-D_v^{-\frac12}HWD_e^{-1}H^TD_v^{-\frac12}$ .

⑧Funcional connectivity network (FCN) matrix each subject:

$\boldsymbol{X}=[\boldsymbol{x}_1,\boldsymbol{x}_2,\ldots,\boldsymbol{x}_N]\in\mathbb{R}^{N\times N}$

where the subscript denotes the order number of each ROI. They set $x_i$ for the feature of $v_i$

⑨Similarity between vertexes:

$\mathrm{Sim}\left(v_i,v_j\right)=e^{-\frac{\left\|\boldsymbol{x}_i-\boldsymbol{x}_j\right\|_2^2}{2\sigma^2}}$

$\sigma$ is set to the mean of the Euclidean distances of all vertex pairs empirically

⑩Followed by ④, they further update incidence matrix to:

$\left.h\left(v_i,e_j\right)=\left\{\begin{array}{ll}\mathrm{Sim}\left(v_i,v_j\right),&\quad v_i\in e_j,\\0,&\quad v_i\notin e_j.\end{array}\right.\right.$

⑪To avoid differences weighted incidence matrix between subjects, they calculated it by average FCN

⑫Convolution operation on $(l+1)$ -th layer:

$\boldsymbol{X}_{(l+1)}=\phi\bigg(\boldsymbol{D}_v^{-\frac{1}{2}}\boldsymbol{H}\boldsymbol{W}\boldsymbol{D}_e^{-1}\boldsymbol{H}^T\boldsymbol{D}_v^{-\frac{1}{2}}\boldsymbol{X}_{(l)}\boldsymbol{\Theta}_{(l)}\bigg)$

where $\boldsymbol{\Theta}_{(l)}$ denotes the learnable weight matrix, $\phi$ represents nonlinear activation function

⑬总的就是说v是ROI，edge感觉是ROI相似度，然后节点特征是FCN的行捏

2.3.2. The proposed CcSi-MHAHGEL

The overall framework:

（1）Multiview hyperedge-aware HGCN for learning embeddings of multiatlas-based FCNs

①They introduced $M$ categories of atlas, then update $X^{(m)}\in\mathbb{R}^{N_m\times N_m}$

②The corresponding convolution:

$\begin{aligned}&\boldsymbol{X}_{(l+1)}^{(m)}\\&=\phi\left(\left(\boldsymbol{D}_{v}^{(m)}\right)^{-\frac{1}{2}}\boldsymbol{H}^{(m)}\boldsymbol{W}_{(l)}^{(m)}\left(\boldsymbol{D}_{e}^{(m)}\right)^{-1}\left(\boldsymbol{H}^{(m)}\right)^{T}\left(\boldsymbol{D}_{v}^{(m)}\right)^{-\frac{1}{2}}\boldsymbol{X}_{(l)}^{(m)}\boldsymbol{\Theta}_{(l)}^{(m)}\right)\end{aligned}$

and they separate and combine the $\boldsymbol{H}^{(m)}$ of patients (PTs) and HCs by:

$H^{(m)}=H_{\mathrm{HC}}^{(m)}\parallel H_{\mathrm{PT}}^{(m)}\in\mathbb{R}^{N_{m}\times2N_{m}}$

③They proposed an adaptive hyperedge weight learning. They first obtain the hyperedge features by:

$\boldsymbol{E}_{(l)}^{(m)}=\left(\boldsymbol{H}^{(m)}\right)^T\boldsymbol{X}_{(l-1)}^{(m)}$

（感觉是个[2e,v] matmul [v,v]的操作，得到个[2e,v]。相当于是一个超边的特征是它连接的所有节点特征的和。但是这样确实一个超边特征是个[1,v]的向量了（⭐但是为什么作者的H是[n,2n]我的是[2e,v]呢我只是想区别一下啦。作者构建了N个超边，所以和N个顶点等同了。每个超边连K个顶点，即从一个ROI开始算距离最近的K-1条边捏））

④To squeez the features, they further:

$\boldsymbol{S}_{(l)}^{(m)}=\mathrm{Sigmoid}\Big(\boldsymbol{E}_{(l)}^{(m)}\boldsymbol{V}_{(l)}^{(m)}\Big)\in\mathbb{R}^{2N_{m}}$

where $\boldsymbol{V}_{(l)}^{(m)}$ denotes learnable vector of weight coefficients. （应该就是[2e,v]mutmal[v,1] ）

⑤The hyperedge weight matrix $\boldsymbol{W}_{(l)}^{(m)}=\mathrm{diag}\Big(\boldsymbol{S}_{(l)}^{(m)}\Big)$ . They mention that the traditional hyperedge weight is not adaptive

⑥The apply max and mean pooling in the last layer of HGCN. For finall output $\boldsymbol{X}_{L}^{(m)}\in\mathbb{R}^{N_{m}\times d_{m}}$ , they concatenate max vector and mean vector to $f^{(m)}\in\mathbb{R}^{2N_{m}}$ . And further concatenating $f$ under different atlases:

$\boldsymbol{f}\in\mathbb{R}^{2\sum_{m=1}^MN_m}=k_1\boldsymbol{f}^{(1)}\parallel k_2\boldsymbol{f}^{(2)}\parallel\cdots\parallel k_M\boldsymbol{f}^{(M)}$

where $k_1,k_2,\ldots,k_M\in\mathbb{R}$ are the weight coefficients which:

$\sum_{i=1}^{M}k_{i}=1,k_i\geq 0,1\leq i\leq M$

（2）Class-consistency module for promoting discrimination across classes in the multiatlas-based FCN embeddings

①To penalize the intra-class dissimilarities and inter-class similarities, they designed:

$\mathcal{L}_{\mathrm{cc}}=\frac{1}{Q}\sum_{y_{i}=y_{j}}\left\|\boldsymbol{f}_{i}-\boldsymbol{f}_{j}\right\|_{2}+\frac{1}{Q}\sum_{y_{i}\neq y_{j}}\mathrm{Max}\big(0,\varepsilon-\|\boldsymbol{f}_{i}-\boldsymbol{f}_{j}\|_{2}\big)$

where $Q$ denotes training subjects（嘛他意思就是同类别的缩小差异，不同类别的扩大差异）

（3）Site-independence module for mitigating site influences in the multiatlas-based FCN embeddings

①The site information $\boldsymbol{s}_i=\left[s_{i,1},s_{i,2},\ldots,s_{i,S}\right]^T\in\mathbb{R}^S$ is calculated by:

$\left.s_{i,p}=\left\{\begin{array}{ll}1,&\quad\text{if subject}i\text{from site}p,\\0,&\quad\text{otherwise}.\end{array}\right.\right.$

where it is one-hot encoder~ and $S$ denotes the type of site

②They introduce the Hilbert–Schmidt independence criterion (HSIC) (the smaller the less dependence) to minimize FCN embeddings $\{f_{i}\}_{i=1}^{Q}$ and the one-hot site vectors $\{s_{i}\}_{i=1}^{Q}$ :

$\mathcal{L}_{\mathrm{si}}=\frac{1}{\left(Q-1\right)^{2}}\operatorname{tr}(\boldsymbol{FUPU})$

where $U\in\mathbb{R}^{Q\times Q}=I-\frac{1}{Q}\mathbf{1}\mathbf{1}^{T}$ , $P_{i,j}=\varphi_{s} (s_{i},s_{j})\in\mathbb{R}^{Q\times Q}$ and $F_{i,j}=\varphi_{f}\left(f_{i},f_{j}\right)$ . Both $\varphi$ are kernel functions, such as linear (the authors choosed), polynomial, and radial basis. Thus:

$F_{i,j}=\boldsymbol{f}_i^T\boldsymbol{f}_j \, \mathrm{and} \, P_{i,j}=\boldsymbol{s}_i^T\boldsymbol{s}_j$

（4）Classification and loss function

①The loss function: $L=L_{ce}+\alpha L_{cc}+\beta L_{si}$ , where $\alpha$ and $\beta$ are tunable hyperparameters or regularization parameters（CE就不多说了吧？交叉熵啊）

2.4. Experimental results

2.4.1. Dataset and data preprocessing

①Dataset: ABIDE I/II（作者说的是”ABIDE“，但是根据有21个站点来推断，应该是把I和II加在一起选的，最大的四个）

②Samples: 167 ASD and 188 HC:

③Preprocessing pipeline: DPARSF

④Preprocessing processes: "head motion correction, co-registration, spatial normalization to the standard Montreal Neurological Institute (MNI) space, spatial smoothing with a 6-mm full width half maximum (FWHM) Gaussian kernel, and band-pass filtering at 0.01 to 0.1Hz."

⑤Atlases: AAL 116 ad HO 111, $X^{(1)}\in\mathbb{R}^{116\times116}$ for AAL and $X^{(2)}\in\mathbb{R}^{111\times111}$ for HO

⑥FCN: "Fisher’s z-transformed Pearson’s correlation coefficient between BOLD time series of each pair of ROIs"

2.4.2. Experimental settings

①Cross validation: 10 fold

②Running times: 5

③Metrics: accuracy, sensitivity, precision, F1-score, and AUC

④Hyperparameters: $K=3$ , hidden dim = 256/64 in the first/second HGCN layer, $\phi$ is ReLU, weight coefficients $k_1=0.3,k_2=0.7$ , MLP: 128/32/2, regularization parameters $\alpha =1 \times 10^{-3}$ , $\beta =5 \times 10^{-3}$

⑤Epoch: 30

⑥Batch size: 24

⑦Learning rate: 0.001

2.4.3. Methods for comparison

①Briefly listing and introducing the parameter settings of compared models. All the models adopt the same convolution layers, MLP layers, hidden dim numbers, pochs, batch size, activation function, and learning rate for fair

2.4.4. Classification performance

①Comparison table:

the multiview methods (i.e., MGEL, MHGEL, and CcSi-MHAHGEL) mostly perform better performance than all those single-view methods (i.e., SVM, DNN, BrainNetCNN, GEL, BrainGNN, GAT and HGEL) and the hypergraph-based deep learning methods (i.e., HGEL, MHGEL, and CcSi-MHAHGEL) also outperform simple-graph-based counterparts (namely GEL and MGEL)

2.4.5. Ablation study

①Ablation study:

where it concludes different signle atlas models, HGEL with HGCN (HAHGEL), MHGEL with HGCN (MHAHGEL), and CcSi-MHAHGEL w/o loss

2.4.6. Construction of the shared incidence matrix

①Ablation between different readout methods:

2.4.7. Most discriminative ROIs discovered by CcSi-MHAHGEL

①They got class activation values of ROIs in fully connection layer by Grad-Cam and visualized the top 10 by BrainNet Viewer:

②They start to analyse some medical findings

visuospatial adj. 视觉空间；视觉空间的

2.5. Discussions

2.5.1. Visualization of FCN embeddings

①2D t-Distributed Stochastic Neighbor Embedding (t-SNE)

2.5.2. Influence of key hyperparameters

①Ablation of $K$ :

②Ablation of $\alpha$ and $\beta$ :

③Ablation of $k_1$ and $k_2$ :

2.5.3. Generalization ability of the class-consistency and site-independence modules

①Adding two loss functions to other models:

2.5.4. Results of CcSi-MHAHGEL with three-atlas-based FCNs

①Three atlas based ablation:

2.5.5. Results of CcSi-MHAHGEL on the whole ABIDE

①They choosed 19 sites which more than 10 subjects with 372 ASD and 433 HC:

2.5.6. Influence of hyperedge construction schemes

①Introducing the impacts

2.5.7. Limitations and future work

①They did not consider sex, age and other influences

②There are substitution of HSIC

③They only consider static graph

2.6. Conclusion

3. Reference List

Wang, W. et al. (2024) 'Multiview hyperedge-aware hypergraph embedding learning for multisite, multiatlas fMRI based functional connectivity network analysis', Medical Image Analysis, 94. doi: Redirecting

夏莉莉iy

关注

30
点赞
踩
29

收藏

觉得还不错? 一键收藏
0
评论
[论文精读]Multiview hyperedge-aware hypergraph embedding learning for multisite, multiatlas fMRI based

计算机-人工智能-脑科学与类脑智能
复制链接

扫一扫