[论文精读]Classification of Brain Disorders in rs-fMRI via Local-to-Global Graph Neural Networks

①The authors emphasize the current fMRI classification models ignores non-imaging information and relationships between subjects or misidentify brain regions or biomarkers

②Then, they put forward a local-to-global graph neural network (LG-GNN) to solve these problems

2.2. Introduction

①The diagnosis of Autism Spectrum Disorder (ASD) and Alzheimer’s disease (AD) is limited and inexperienced

②A Harvard-Oxford atlas maps of brain

③Graph neural networks (GNNs) has been found suitable for brain network analysis

④There are two types of GNN: regional brain graph and subject graph. The first one is good at local brain regions and biomarkers analysis, but ignores age, gender or relationships between subjects. And subject graph is just the opposite

⑤Therefore, they combine two methods to combine their advantages. They first adopt local, then expand to global

⑥Contribution: a) end to end LG-GNN, b) a pooling strategy based on an attention mechanism, c) excellent performances in 2 datasets

etiological adj.病因;病原学的，病原学的

aberrant adj.异常的;反常的;违反常规的

2.3. Related Work

2.3.1. Handcrafted Methods

①Functional connection construction by calculating Pearson correlation coefficient or "extracting weighted local clustering coefficients from the brain connectivity network from rs-fMRI, and then employed multiple-kernel based-SVM algorithm for the subsequent MCI classification"

②"Using SVM to classify AD from normal control (NC) after PCA and Student’s t-test for dimension reduction based on shape and diffusion tensor imaging"

③Also, different ROI may impact on the prediction accuracy

2.3.2. GNN-Based Methods

（1）GNNs Based on Regional Brain Graphs

Models such as DS-GCNs, s-GCN, MVS-GCN and mutual multi-scale triplet GCN etc. adopt regional graphs

（2）GNNs Based on Subject Graphs

Models such as GCN, InceptionGCN and SAC-GCN etc. adopt subject graphs

2.4. Proposed Method

①Framework of LG-GNN:

2.4.1. Local ROI-GNN

（1）Regional Brain Graph Construction

Approach of graph $G_{local}=\{\mathbf{V},\mathbf{A}\}$ construction, each graph is represented as $\mathbf{X}=\left[x_1,\ldots,x_n\right]^\top$ :

$\left.\mathbf{A}_{ij}=\left\{\begin{array}{ll}\frac{Cov(v_i,v_j)}{\sigma_{v_i}\sigma_{v_j}},&\text{if }v_i\text{and }v_j\text{are adjacent},\\1,&\text{if }i=j,\\0,&\text{otherwise}.\end{array}\right.\right.$

（2）Local ROI-GNN Model

①Local ROI-GNN framework:

it consists of three graph convolution (GC) layers and the Self-Attention Based Pooling (SABP) module

②The activation in GC layer:

$\begin{aligned} \mathbf{X}^{(t)}=& \mathrm{ReLU}(\mathrm{GC}(\mathbf{X},\mathbf{A}) \\ =& \operatorname{ReLU}\left(\mathbf{D}^{-\frac12}\mathbf{AD}^{-\frac12}\mathbf{X}^{(t-1)}\mathbf{W}^{(t)}\right) \end{aligned}$

where $\mathbf{D}=\mathrm{diag}\left(\sum_j\mathbf{A}_{1j},\sum_j\mathbf{A}_{2j},\ldots,\sum_j\mathbf{A}_{nj}\right)$ denotes the degree matrix of $\mathbf{A}$ ;

$\mathbf{W}^{(t)}$ represents trainable weight matrix in the $t$ -th layer;

$\mathbf{X}^{(t)}=\left[x_1^{(t)},x_2^{(t)},\ldots,x_n^{(t)}\right]^\top$ is node representation.

③Receptive field: one hop neighborhoods

④For getting two-hops information, they adopt 2 GC layers. Then they get $\mathbf{X}^{(2)}\in\mathbb{R}^{n\times c}$

（3）Self-Attention Based Pooling

①Pooling is essential in preserving and highlighting important ROI

②Transform $\mathbf{X}^{(2)}\in\mathbb{R}^{n\times c}$ to attention score $\textbf{z}$ by $\mathbf{z}=\mathrm{GC}^{\prime}(\mathbf{X}^{(2)},\mathbf{A})$ .

③Through $\tilde{\mathbf{z}}=\tanh(\mathbf{z})$ , selecting the top- $k$ nodes $\mathrm{idx=}\mathrm{topk}(\tilde{\mathbf{z}},k)$

④Then, follows a pooling layer (?) $\hat{\mathbf{X}}^{(2)}=\mathbf{X}^{(2)}(\mathrm{idx},:)$ and a $\tilde{\mathbf{X}}^{(2)}=\hat{\mathbf{X}}^{(2)}\odot\tilde{\mathbf{z}}$

⑤Finally, obtain new adjacency matrix $\widetilde{\mathbf{A}}=\mathbf{A}(\mathrm{idx},\mathrm{idx})$

⑥They designed a loss to separate ROI weights

$\begin{gathered} \mathcal{L}_{MI}= \frac1n\sum_{i=1}^n\mathrm{MLP}(\mathrm{Concat}(\mathrm{GC}(\mathbf{X}^{(2)},\mathbf{A}),\mathbf{X}^{(2)}) \\ -\log\frac1n\sum_{i=1}^ne^{\mathrm{MLP}(\mathrm{Concat}(\overline{\mathbf{X}},\mathbf{X}^{(2)}))} \end{gathered}$ $m$

where $\overline{\mathbf{X}}$ is randomly shuffled $\mathbf{X}$ in channels （我也不知道有什么用啊）.

⑦The output will be:

$\mathbf{Y}=\mathrm{GC}(\tilde{\mathbf{X}}^{(2)},\tilde{\mathbf{A}})$

where $\mathbf{Y}=[y_1,y_2,\ldots,y_k]^{{\top}}$

⑧我很想知道为什么作者说SABP可以考虑到拓扑关系啊，脑图不是无向图吗而且也不会有什么先后关系吧？？

2.4.2. Global Subject-GNN

（1）Subject Graph Construction

①There are $m$ subjects, i.e. $m$ vector $\textbf{V}{}'$

② $\textbf{A}{}'=\textbf{C*W}$ is the adjacency matrix of $\textbf{V}{}'$ , where $\textbf{C}$ is a binarized connectivity matrix of the combination of non-image information and information (with values greater than 0.4 becoming 1, and values less than 0.4 becoming 0)

③There is a similarity matrix $\mathbf{S}_1=\exp\left(-\frac{\left[\rho(\mathbf{Y}_i,\mathbf{Y}_j)\right]^2}{2\sigma^2}\right)$ , where $\rho$ denotes the correlation distance, $\sigma$ is mean of $[\rho(\mathbf{Y}_i,\mathbf{Y}_j)]^2$

④Node similarity metric $\textbf{S}_{2}$ is constructed by non-image information

⑤ $\mathbf{C}^{\prime}=\mathbf{S}_1\odot \mathbf{S}_2$

⑥Weight matrix $\mathbf{W}_{ij}=\frac{\mathrm{Sim}(\mathrm{MLP}(\eta_i),\mathrm{MLP}(\eta_j))+1}2$ , where $\eta _{i}$ and $\eta _{j}$ are both non-image information, Sim reperesents the cosine similarity, the 2 MLP layers share the same weights

（2）Global Subject-GNN Model

①A multi-scaled residual model was proposed as:

②Cheb conv is:

$\mathbf{g}_\theta\star h=\mathbf{U}\mathbf{g}_\theta(\mathbf{\Lambda})\mathbf{U}^\top h$

where $h\in\mathbb{R}^m$ looks like the eigenvalues of each node;

$\mathbf{g}_\theta=\operatorname{diag}(\theta)$ denotes a filter;

$\theta\in\mathbb{R}^m$ denotes parameter;

$\star$ denotes convolution operator;

$\textbf{U}$ is come from Laplace matrix $\mathbf{L}=\mathbf{I}-\mathbf{D}^{-\frac12}\mathbf{A}\mathbf{D}^{-\frac12}=\mathbf{U}\mathbf{\Lambda}\mathbf{U}^\top$ ;

③To save the calculating time, they approximate the above function to:

$\mathbf{g}_\theta\star h\approx\sum_{k=0}^{K-1}\theta_kT_k(\tilde{\mathbf{L}})h$

where $\tilde{\mathbf{L}}$ is a rescaled Laplace graph, $\theta_k$ denotes a learnable parameter

④The recursion might be $T_k(\tilde{\mathbf{L}})=2\tilde{\mathbf{L}}T_{k-1}(\tilde{\mathbf{L}})-T_{k-2}(\tilde{\mathbf{L}})$ , $T_0(\tilde{\mathbf{L}})=1$ , $T_1(\tilde{\mathbf{L}})=\tilde{\mathbf{L}}$

⑤Given the $\mathbf{H}=\left[Y_1,Y_2,\ldots,Y_m\right]^\top$ , there is $\mathbf{H}^{(l+1)}=\sum_{k=0}^{K-1}\theta_k^{(l)}T_k(\tilde{\mathbf{L}})\mathbf{H}^{(l)}$

⑥The final embedding $\mathbf{Z}=\sum_lw^{(l)}\odot\mathbf{H}^{(l)}$ ,

where $w$ is learnable weight $w^{(l)}=\text{Softmax}(r^{(l)})=\frac{\exp(r^{(l)})}{\sum_l\exp(r^{(l)})}$ ;

$r$ denotes random initialized learnable weight;

（3）Total Training Loss

①They adopt cross entropy loss as the global loss:

$\mathcal{L}_{CE}=\frac1N\sum_i-[q_i\cdot\log(p_i)+(1-q_i)\cdot\log(1-p_i)]$

②They combine local $\mathcal{L}_{MI}$ and global $\mathcal{L}_{CE}$ to get total loss:

$\mathcal{L}_{total}=\mathcal{L}_{CE}+\lambda\mathcal{L}_{MI}$

where the hyper parameter $\lambda$ is set as 0.1

2.5. Materials and Classification Evaluation

2.5.1. Materials

They choose two datasets ABIDE, ADNI in 4 tasks:

（1）ABIDE Dataset

①They only select 871 samples from 1112 subjects with 403 ASD and 468 NC in 20 different sites.

②Preprocess pipeline: C-PAC

③Space: normalized in MNI152 space

（2）ADNI Dataset (NC and MCI)

①Preprocess: standard protocol

②Excluding significant artifacts or head movements beyond 2mm amplitude

③They choose 134 subjects with 96 MCI and 40 AD （不是，我有点不太理解这句话，40+96也不等于134吧？）

（3）ANDI Dataset (pMCI and sMCI)

①pMCI patients: deteriorating within 36 months; sMCI patients: do not deteriorate

②Atlas: Harvard-Oxford

③They choose 41 pMCI and 80 sMCI subjects excluding NaN data

④Preprocessing: standard procedures on GRETNA toolkit

2.5.2. Experimental Setting

①Optimizer: Adam

（1）Parameters of ABIDE

①Dropout rate: 0.3

②Learning rate: 0.01

③Maximum epochs: 400

④Non-image data: acquisition site, gender

（2）Parameters of ANDI

①Dropout rate:0.3

②Learning rate: 0.01

③Maximum epochs: 300

④Non-image data: gender and age

②Chebyshev polynomial order K: 3

③Cross-validation: 10-fold, 9 for training and 1 for test

④Evaluation metrics: classification accuracy (Acc), area under the curve (AUC), sensitivity (Sen) and F1-score

⑤这里的ROI Dimension是什么东西？为什么是2000，2140之类的？Non-imaging dimension denotes (number of subjects, binary classification)

2.6. Experimental Results

2.6.1. Classification Results on Different Datasets

（1）Classification Results on ABIDE

①Comparison with handcrafted methods: Ridge Classifier, SVM, Random Forest classifier

②Comparison with GNN-based methods: GCN, GAT, BrainGNN, MVS-GCN, PopulationGNN, InceptionGCN, EV-GCN, Hi-GCN

③Comparison with deep neural network (DNN) based methods

④The classification results (NI denotes whether adopting non-image data or not):

（2）Classification Results on ADNI (NC and MCI)

①Models in ANDI mostly outperform those in ABIDE in that data in ABIDE is highly heterogeneous (they come from different acquisition site)

②The classification results (NC and AD):

③The classification results on NC and MCI:

④The classification results on pMCI and sMCI:

2.6.2. Ablation Studies

（1）Ablation Study for Local ROI-GNN

①They compared local ROI-GNN with GAT, GIN, GraphSAGE, ChebNet, GCN, which means adopting these SOTA models in their model (replace local ROI-GNN)

②They found SABP module significantly enhances the performance. Specifically, they reckon appropriate pooling is vital. Excessive pooling may retain a lot of noise, but too small pooling may cause the topology information of the graph to be compressed. Thus, they choose 0.9 pooling rate.

③Replaced classification results:

④Different pooling rate in SABP module:

（2）Ablation Study for Global Subject-GNN

①They replaced subject-GNN by GATConv, GINConv, GraphSAGEConv, GCNConv

②Replaced classification results:

③To evaluate the effectiveness of AWAB module, they tried with or without (adopting output of the last Cheb block as the final output) AWAB in the table above

2.6.3. Biomarker Detection

①They obtain weights of ROI through SABP module, and select top 10 ROI as biomarker

②Mutual information loss $\mathcal{L}_{MI}$ is what they distinguish ROIs with putting important ROI weights to 1, other to 0

③The top 10 ROIs with the greatest impact on autism and Alzheimer's disease:

2.6.4. Discussion

①Acquisition site significantly affects the results, and gender impacts on classification as well in ABIDE dataset:

②Gender impacts more than age in ANDI:

③Other non-imaging data can also be used, such as IQ and genetic informatio（？？？你可以再写离谱一点吗我真的，这）

quotient n.商(除法所得的结果)

2.7. Conclusion

Their model LG-GNN with local and global modules achieve excellent performance

3. Reference List

Zhang, H. et al. (2022) 'Classification of Brain Disorders in rs-fMRI via Local-to-Global Graph Neural Networks', IEEE, vol. 42 (2), pp. 444-455. doi: 10.1109/TMI.2022.3219260