[论文精读]D-MHGCN: An End-to-End Individual Behavioral Prediction Model Using Dual Multi-Hop Graph--CSDN博客

①The graph is represented by $G^{C}=(V^{C},A^{C},X^{C})$ , where $V^C$ denotes the cortical vertices of the brain, $A^C$ is the spatial proximity of vertices in T1w images, $X^C$ is FC matrix come from fMRI data

②Cortical parcellation separately applied on two brain hemispheres

③Connectivity weight in $A^C$ : the inverse of their geometric distance. "邻接矩阵A^C中的边表示两个顶点是否在皮质表面共享一个三角形区域"（0.0我怎么知道共享不共享？作者后来又说“该图显示了稀疏但高度局部化的空间连接，其中每个顶点平均连接到大约六个最近的邻居”）

④ $\left | V^C_{left} \right |=2392$ and $\left | V^C_{right} \right |=2332$

⑤Construction of FC: Pearson correlation between BOLD signals, all the negative value is set to 0, and only remains the connections which the spatial distance was less than 1

3.3.2. Individual Functional Connectivity Network Construction

（1）Problem Statement

①For $n$ subject, the cortical graph is $G_{n}^{C}=(V_{n}^{C},A_{n}^{C},X_{n}^{C})$ , $V_{n}^{C}$ denotes vertex set, $A_{n}^{C}$ represents adjacency matrix, $X_{n}^{C}$ is FC matrix

②The region level $G^{FC}_n$ is generated by $G_{n}^{C}$ and fMRI time series

（2）Multi-Hop Graph Convolutional Network

①Backbone: GCN

②Multi-hop aggregation method:

they defined the hop set

③The node feature matrix in the $l$ -th layer:

$(X_n^C)^l\in R^{|V_n^C|\times d_l}$

④Multi-hop convolution:

$(X_n^C)^{(l+1)}=\prod\limits_{0\le k\le K}\sigma((\tilde{A}_n^C)^k(X_n^C)^lW_k^l)$

where $\tilde{A}_n^C=(\tilde{D}_n^C)^{-1/2}(A_n^C+I)(\tilde{D}_n^C)^{-1/2}$ , $\sigma$ denotes ReLU activation function

⑤The final node feature is represented by:

$f_n^C=X_n^CO$

where $O$ denotes the trainable parameters

⑥Classifier: SoftMax

toy example 简单的示例

（3）Individual Cortical Parcellation

①Atlas: Schaefer-100 atlas and the Human Brainnetome Atlas (BNA)

②By mapping to the standard space, they got the label vector $y_i$ for each vertex $i$

③Employed cross-entropy loss $L_{par}$ to calculate the difference between the actual regional labels $y_i$ (from group-level template) and the predicted regional probabilities $p_i$ (obtained from multi-hop GCN) with:

$L_{par}=\sum_{i=1}^{|V_n^C|}\sum_{j=1}^Jy_i^j\log(p_i^j)$

where $J$ denotes the number of regions in predefined group-level template（怎么给到的每个顶点预测概率？）, $p_i^j$ denotes the probability that node $i$ belongs to region $j$ , $y_i^j=1\, or\, 0$ is whether node $i$ is in region $j$ in the templet

④ $J$ in Schaefer and BNA is 50 and 106

（4）Individual Functional Connectivity Estimation

①The region level FCN: $G^{FC}=\left ( V^{FC},A^{FC} \right )$

②Dividing the brain of each subject to $\left | V^{FC}_n \right |$ regions and calculating the PC of each region with all negative values are 0. The generated FC is regarded as the $A^{FC}_n$ in $G^{FC}_n$

3.3.3. Individual Behavior Score Estimation

（1）Multi-Hop Graph Convolutional Layer

①The FC feature in the $l$ -th layer:

$(X_n^{FC})^l\in R^{|V_n^{FC}|\times d_l}$

②The FC conv layer:

$(X_n^{FC})^{(l+1)}=\underset{0\leq k\leq K}{\parallel}\sigma\left((\widetilde{A}_n^{FC})^k(X_n^{FC})^lU_k^l\right)$

where $\tilde{A}_{n}^{FC}=\left(\tilde{D}_{n}^{FC}\right)^{-1/2}(A_{n}^{FC}+I)\left(\tilde{D}_{n}^{FC}\right)^{-1/2}$

（2）Pooling Layer

①Aggregation to get feature vector for each subject:

$f_{n}^{FC}=\frac{1}{|V_{n}^{FC}|}\sum_{v=1}^{|V_{n}^{FC}|}(x_{v})_{n}^{FC}$

...就是一个平均池化嘛

（3）Behavior Score Assessment

①The scores are calculated by fully connected layer:

$\hat{z}_n=f_n^{FC}Q$

where $Q$ denotes the trainable parameters

②Measurements: MAE:

$L_{net}=\frac{1}{N}\sum_{n=1}^{N}E\left(z_{n},\hat{z}_{n}\right)$

where $z_n$ is the groud truth and $\hat{z}_n$ denotes the predicted score

3.3.4. Total Loss

①The total loss:

$L_{\mathrm{total}}=\alpha L_{par}+L_{net}$

3.4. Experimental Settings

3.4.1. Dataset

（1）Data Information

①Dataset: HCP S1200

②State of subjects: awaken, eyes open and staring a bright crosshair projected against a dark background

③Sample: 910

（2）Imaging Data Preprocessing

①Preprocessing of sMRI: 1) distortion correction, 2) alignment of T1w and T2w images, 3) bias field correction, 4) cortical tissue segmentation (gray matter, white matter, and cerebrospinal fluid), and 5) cortical reconstruction, 6) resampling

②Tool: FreeSurfer

③Preprocessing of fMRI: 1) gradient distortion correction, 2) head motion correction, 3) EPI distortion correction, 4) registration to Montreal Neurological Institute (MNI) space, 5) intensity normalization to the global mean, and 6) removal of non-brain voxels， 7) mapping, 8) resampling

（3）Behavioral Scores

①Cognitive tests: executive function-related test (Cognitive Flexibility), languagerelated test (Story Difficulty Level), and comprehensive cognitive test (Fluid Intelligence)

②Evaluation methods: NIH Toolbox

3.4.2. Implementation Details

①Learning rate: 0.005

②Optimizer: Adam

③L2 Norm: coefficient of 5e–5

④Epoch: pretraining 30 for individual functional connectivity network construction, and train individual behavior score estimation 30 rounds. Lastly, jointly train they two with 20 rounds

⑤Layers of NN: 7, {32,64,50,32,64,64,1} in Schaefer100 and {32,64,106,32,64,64,1} in BNA

3.4.3. Evaluation Metrics

（1）Behavioral Prediction

①Evaluation method: "The accuracy of behavioral prediction was evaluated by calculating the Pearson correlation between the actual behavior scores and the assessed behavior scores. The range of correlation is from −1 to 1, with higher values indicating a better behavioral prediction result."

（2）Cortical Parcellation Quality

①Evaluation: Silhouette Width:

$\begin{aligned}SI(M)=\frac{1}{J}\sum_{j=1}^{J}\frac{a_j-b_j}{\max{\{a_j,b_j\}}}\end{aligned}$

where $a_j$ denotes the average FC among nodes within the parcel $j$ , $b_j$ is the average FC of nodes in parcel $j$ cross parcels

3.5. Experimental Results and Analysis

3.5.1. Parameter Analysis

①Parameter ablation on 5 fold cross validation:

3.5.2. Comparison With State-of-The-Art Methods

（1）Behavioral Prediction Accuarcy

①Performance table:

（2）Individual Cortical Parcellation Quality

①Sample: 910 subjects with 728 for training and 182 for testing

②Comparison figures:

3.5.3. Effectiveness of the Proposed Behavioral Prediction Framework

①Module ablation:

3.5.4. Effectiveness of Multi-Hop Graph Convolutional Operator

①Hop ablation:

3.5.5. Analysis of Individual Cortical Parcellation

①Analysis of individual cortical parcellation:

3.5.6. Discussion of Future Clinical Applications

①Personalized cortical segmentation map is important

3.6. Conclusion

5. Reference

Wen, X. et al. (2024) 'D-MHGCN: An End-to-End Individual Behavioral Prediction Model Using Dual Multi-Hop Graph Convolutional Network', IEEE Journal of Biomedical and Health Informatics, 28(10): 6130-6140. doi: 10.1109/JBHI.2024.3420134