[论文精读]A New Brain Network Construction Paradigm for Brain Disorder Via Diffusion-Based Graph Contra--优快云博客

①Introduced the datasets used, commonly used imaging modalities for constructing brain networks, and the earliest methods for constructing brain networks

2.3.2. Graph neural network methods

①Listing GNNs methods and claiming that they heavily rely on data preprocessing

2.3.3. Deep learning methods

①Listed methods for automatically learning brain network features, processing preprocessing steps, etc

2.4. Methodology

①Overall framework:

2.4.1. Brain Region-Aware Module

（1）Brain region feature extraction and alignment

①Brain atlas: AAL, and align to the standard regions:

$\psi^*=\arg\min_\psi\mathcal{L}_{sim}(f,m\circ\psi)+\mathcal{L}_{reg}(\psi)$

where $\psi^*$ is the best registration field, $m$ denotes the moving image, $\mathcal{L}_{sim}$ denotes the similarity function, $\mathcal{L}_{reg}$ denotes the regularization penalty for the registration field

（2）The structure of proposed BRAM

①They employed Trilinear interpolation in 3D DTI registration on moving image $m$ :

$m'=m\circ\psi$

②"The final output of the first stage model is:"（啊？？？为什么这里不做介绍啊？denoising dual U-Net and deformation network为什么也没展开介绍呢？）

$M_i= m'\cdot \text{mask}(region_i)$

deformation n.变形；畸形；残废；变丑；破相；损形

（3）Reverse alignment process

①They aim to learn a conditional score function of the deformation between moving image $m$ and fixed reference image $f$ :

$c=\left ( m,f \right )$

②The fixed image can be regard to a target: $x_0=f$

③Loss function of diffusion network:

$\mathcal{L}_{\mathrm{diffusion}}\left(c,x_{t},t\right)=\mathbb{E}_{\epsilon,x_{t},t}\left\|D_{\theta}\left(c,x_{t},t\right)-\epsilon\right\|_{2}^{2}$

where $D_{\theta }$ is the diffusion network

④Optimization objective of diffusion model:

$\min_{D_\theta,M_\psi}\mathcal{L}_\text{diffusion }(c,x_t,t)+\lambda\mathcal{L}_\text{regist }(m,f)$

where $\mathcal{L}_\text{regist }$ denotes registration loss, $\mathcal{L}_\text{regist }$ contains $\mathcal{L}_\text{sim}$ and $\mathcal{L}_\text{smoo}$ , $M$ denotes input moving images, $F$ denotes input reference images

⑤Optimization objective of registration model:

$\psi=\arg\min\mathcal{L}_{sim} (F,M\circ\psi)+\mu\mathcal{L}_{\mathrm{smoo}} (\psi)$

where $\mathcal{L}_{sim}$ denotes the image structure similarity loss, $\mathcal{L}_{smoo}$ denotes the smoothness constraint on the gradient of the deformation field

⑥Aligned image generation algorithm:

2.4.2. Brain Network Optimization Module

①The workflow of reconstructing the embedding features of brain network nodes:

②Generating diffusion map by thermonuclear features:

$S^\text{heat}=\exp\left(tAD^{-1}-t\right)$

where $t$ denotes the diffusion time, and it slightly aggregates neighbor nodes

③Loss of constrastive learning:

$\mathcal{L}_{\mathrm{cont}}=\max_{\theta,\omega,\phi,\psi}\frac{1}{|\mathcal{G}|}\sum_{g\in G}\sum_{i=1}^{|g|}\Big[\mathrm{MI}(h_{i}^{\alpha},h_{g}^{\beta})+\mathrm{MI}(h_{i}^{\beta},h_{g}^{\alpha})\Big]$

where $MI$ denotes the mutual information function （这里对应图中间两个蓝框框，是图嵌入和节点嵌入都要算点积的）

④Normalizating embedding matrix by inner product and constraining them to 0-1 by activation function（这个在图里又没有了....要回到最初大图才有）:

$\mathcal{L}_{\mathrm{ref}}\left(\mathcal{G}\right)=-\frac{1}{90\times90}\sum_{i,j}(A_{p}[i,j]-A_{\mathrm{ref}}[i,j])^{2}$

where $A_{p}[i,j]=\sigma (h_{i}^{\top}h_{j})$

2.4.3. Classifier Design and Model Training

①Classifier: consist of 3-layer MLP, 2 ReLU, and Softmax with 4 neurons

②Multiclassification loss:

$\mathcal{L}_{\mathrm{cls}}\left(p\right)=-\sum_{i=0}^{C-1}y_i\log(p_i)+(1-y_i)\log(1-p_i)$

③The total loss of DGCL:

$\mathcal{L}_{total}=\mathcal{L}_{\mathrm{cont}}+\mathcal{L}_{\mathrm{cls}}\left(p\right)+\alpha\mathcal{L}_{\mathrm{ref}\left(\mathcal{G}\right)}$

④Training process:

2.5. Experiments

2.5.1. Datasets and Preprocessing

（1）AD

①Dataset: ADNI

②Modality: DTI

③Classes: normal control group (NC), early mild cognitive impairment (EMCI), late mild cognitive impairment (LMCI), and AD

④Statistic information:

（2）ASD

①Datasets: San Diego University, NYU Langone Medical Center, and Trinity Centre for Health Sciences State

②Statistic information（作者标号写错了，这是table 2）:

（3）Preprocessing

①Tool: PANDA, to get reference structural brain network matrix

②Preprocessing steps: "converting the initial DICOM format of the data to NIFTI format, skull stripping, fiber bundle resampling, and head motion correction. Then we calculate the fractional anisotropy (FA) coefficients by fitting the tensor model using the least squares method, and output the DTI data. After resampling, all subjects have 91×109×91 voxels in DTI, with a voxel size of 2mm × 2mm × 2mm."（所以并没有完全取代好吧！？有什么用！！就只是帮助register是吧）

③Then register the image to get structural connectivity $\hat{A}$

④Atlas: AAL 90

2.5.2. Experiment Settings

①Optimizer: Adam

②Initial learning rate: 0.0001 with weight decay（没说咋decay）

③Epoch: 300

④Batch size: 16

⑤Cross validation: 5 fold, 4 for training and 1 for testing

⑥Metrics: ACC, SEN, SPE and AUROC

⑦Initialization scheme of parameters: Xavier

2.5.3. Ablation Study and Analysis

（1）Hyperparameters Analysis

①Ablation of loss coefficient $\lambda$ of the alignment network and constraint coefficient $\mu$ of the smoothness of the deformation field gradient（作者说 $\lambda =2$ 是原论文的最优值，因此他们也选2了，只调试 $\mu$ ）:

②Ablation study of coefficient $\alpha$ of the reference brain network:

③Visualization under different $\alpha$ :

④Module ablation:

（2）Loss functions Analysis

①Loss ablation:

2.6. Discussion

2.6.1. Compasirons with Baseline

①Comparison on AD（就比这俩？SOTA呢）:

②Comparison on ASD:

③t-SNE visualization on ADNI and ABIDE:

④"comprehensive comparison"（？？？？我也没见得有多综合）:

2.6.2. The optimized Brain Networks

①Four binary classification on ADNI:

②The changes in brain connectivity from NC to EMCI, LMCI, and AD:

③Connection strength on different stages of AD:

④Connectivity difference between HC and ASD:

where these matrices are obtained by averaging weight scores on correctly classified test data（加权是咋加？）

①The calculation method of degree centrality:

$C_{deg}\left(v_i\right)=\frac{\sum_jA[i,j]}{\sum_j\mathbf{1}(A[i,j]>0)}$

②AD related ROIs indentified by PANDA, DGCL and overlap region:

③Specific important regins:

④8 subnetwork mask ablation on ABIDE:

⑤The top 5 important regions of ASD:

2.7. Conclusion

①They aim to expand this work to other disease

②Limitations of their work: do not reveal the pathogenesis, small sample size

3. Reference

Zong, Y. et al. (2024) 'A New Brain Network Construction Paradigm for Brain Disorder Via Diffusion-Based Graph Contrastive Learning', IEEE Transactions on Pattern Analysis and Machine Intelligence, 1-16. doi: 10.1109/TPAMI.2024.3442811