[IEEE TMI 2024]Spatio-Temporal Graph Hubness Propagation Model for Dynamic Brain Network Classificat-CSDN博客

①(a) traditional methods separately extract spatial and temporal feature; (b) some dynamic methods extract feature by sliding windows, but they may ignore the relevant between different time window; (c) their model learn the graph hubness of each brain region with correlation and dynamic evolution by optimal transport:

2.3. Related Work

2.3.1. Dynamics in Brain Network

①"They only studied the independent changes of the brain network within each time window, which ignores the evolution process of functional brain networks."（其实没太get到作者这句话，的确现有的直接划分时间窗然后提取特征看上去比较独立。但进化过程emm，为什么就重要呢？类比信号改变的结果和改变的趋势，emmm。就好像这个作者是在求导一样！）

2.3.2. Graph Convolutional Network

①就介绍图卷积啥的，但也是很基础的GCN而不是什么特别的模型

2.4. Proposed Method

①The pipeline of proposed method:

2.4.1. Network Dynamics Model of Functional Brain Network

①Original rs-fMRI time-series data: $X=(x_1,x_2,\cdots,x_N)^T\in R^{N\times M}$ with $N$ brain regions and $M$ time points

②Dynamic functional brain network (DFBN) construction: divide $X$ to $T$ non-overlapped windows and calculate Pearson Correlation (PC) for each window with connection threshold $\alpha$ .

③Node importance: measured by PageRank

④The hubness of node:

$h(u)_{i+1}=\sum_{v\in B(u)}\frac{h(v)_i}{D_v}$

where $B\left(u\right)$ denotes the neighbors of node $u$ , $D_{v}$ denotes the degree of node $v$ , $h\left ( v \right )_i$ denotes the hubness scores of node $v$ at the $i$ -th iteration. The initial hubness value of each node is $\frac{1}{N}$

⑤Transmission process, when $t-1^{th}$ subnetwork is source domain, then the $t^{th}$ is regarded as the target domain:

（win1这个表示真的很搞笑...虽然我知道是window 1）

⑥Hubness between adjacent nodes is learned by Kantorovich discrete optimal transport:

$\begin{aligned} \min_{f_{t-1}\in R^{N\times N}}\langle W_{s},f_{t-1}\rangle_{F}, \\ s.t.f_{t-1}\mathbf{1}_{N} & =h_{t-1},\quad f_{t-1}{}^T\mathbf{1}_N=h_t, \\ f_{t-1}\mathbf{1}_{N} & =\Sigma_jf_{t-1_{ij}},\quad f_{t-1}{}^T\mathbf{1}_N=\Sigma_if_{t-1_{ij}} \end{aligned}$

where $f_{t-1}$ denotes optimal transport matrix, $W_s$ denotes the transport cost matrix, $\langle\cdot,\cdot\rangle_{F}$ denotes the Frobenius inner product of matrices, $\langle W_s,f_{t-1}\rangle_F=\sum_i\sum_jW_{s_{ij}}f_{t-1_{ij}}$ , $\mathbf{1}_{N}$ is a vecoter with $N$ dimension which all elements are 1

⑦Entropy regularization function:

$\left.H\left(f_{t-1}\right)=-\sum_{i,j}f_{t-1_{ij}}(log\left(f_{t-1_{ij}}\right)-1\right)$

After regularization, the optimization function will be:

$\min_{f_{t-1}\in R^{N\times N}}\langle W_{s},f_{t-1}\rangle_{F}-\varepsilon H(f_{t-1}),\\s.t.f_{t-1}\mathbf{1}_{N}=h_{t-1},{f_{t-1}}^{T}\mathbf{1}_{N}=h_{t}$

where $\varepsilon$ denotes regularization factor

⑧The optimal value is calculated by the matrix deflation:

$f_{t-1_{ij}}=diag(u_i)K_{ij}diag(v_j)$

where $K_{ij}=exp(-W_{s_{ij}}/\varepsilon)$ and $u\circ(Kv)=h_{t-1},\nu\circ(K^Tu)=h_t$ , $\circ$ is element wise product

⑨Iteration process:

$u^{\mathrm{L}+1}=\frac{h_{t-1}}{Kv^\ell},v^{\mathrm{L}+1}=\frac{h_t}{K^Tu^{\mathrm{L}+1}},v^0=1_N$

which gets the results:

$F=(f_1,f_2,\cdots,f_{T-1})^T\in R^{(T-1)\times N\times N}$

2.4.2. Multi-Channel Spatio-Temporal Graph Convolutional Network

①Spatial attention mechanism:

$P=softmax(V_p\cdot Sigmoid((F^TZ_1)Z_2(Z_3F^T)^T+b_p))$

where 除了 $F$ 都是可学习参数

②Temporal attention:

$Q=softmax(V_q\cdot Sigmoid((FM_1)M_2(M_3F^T)+b_q))$

同上

③Graph convolution:

$g_\Theta*_Gx=g_\Theta(L)x=\sum_k^{K-1}\Theta_kT_k(\tilde{L})x$

④CNN:

$\hat{F}^{(l)}=CNN(Relu(g_\Theta*_G(CNN(\hat{F}^{(l-1)}))))$

⑤Algorithm:

2.5. Experiment

2.5.1. Materials

（1）Epilepsy subjects were collected from Jinling Hospital, Nanjing University School of Medicine

①Subjects: 306, with 114 NC, 103 frontal lobe epilepsy (FLE) and 89 temporal lobe epilepsy (TLE)

②Modalities: rs-fMRI and DTI

（2）ADNI

①Collection: ADNI2 and ADNI3

②Subjects: 82NC, 76 significant memory concern (SMC), 69 early mild cognitive impairment (EMCI)

（3）Pre-processing (for two datasets)

①Brain atlas: AAL 90

②Steps: 去原文看吧

2.5.2. Comparison Methods

①Compared methods（不仅有文字介绍还有专门的表...）（诶不对但怎么比了这么多方法）:

2.5.3. Implementation

①Optimizer: Adam with 5e-3 learning rate, 1e-3 weight decay

②Epoch: 300

③Batch size: 20

④CNN: kernel size is 3 with stride 1

⑤GNN: 图卷积核的大小是45！？这啥玩意儿，可能我没用过这么原始的

⑥Cross-validation: 10 fold （但将训练集中 20% 的样本划分到验证集中，并选择验证集上表现最好的模型进行测试。该实验重复 10 次，并报告这 10 个实验的平均结果。）

⑦Hyper-parameter: grid search（啊啊啊啊啊啊啊做期刊的都很有时间啊）

2.5.4. Classification Performance

①Comparison table, with methods which have different brain construction method:

②Comparison table, with methods which have different feature extraction method:

③AUC table:

2.6. Discussion

2.6.1. Analysis of Parameter Sensitivity

①Sensitivity of hyperparameters $\alpha$ and $\varepsilon$ :

2.6.2. Visualization of Dynamic Brain Networks

①作者提出的方法可以捕获变化过程：

2.6.3. Ablation Study

①Module ablation:

2.6.4. Discriminative Brain Connectivity and Regions

①Hubnesses of patients change more than NC

②Significant connections:

③The average values of the top 3 most discriminative dynamic connections of different groups on the four diagnosis tasks:

④The top 10 significant brain regions:

2.6.5. Computational Efficiency Analysis

①Comparison table of computation cost: