[论文精读]Deep Graph Normalizer: A Geometric Deep Learning Approach for Estimating Connectional Brain

最新推荐文章于 2025-04-02 19:15:55 发布

夏莉莉iy

最新推荐文章于 2025-04-02 19:15:55 发布

阅读量1k

点赞数 25

分类专栏：论文精读文章标签：深度学习人工智能机器学习图像处理数据分析神经网络 python

本文链接：https://blog.csdn.net/Sherlily/article/details/142031667

版权

论文精读专栏收录该内容

166 篇文章

订阅专栏

论文全名：Deep Graph Normalizer: A Geometric Deep Learning Approach for Estimating Connectional Brain Templates

论文代码：GitHub - basiralab/DGN: How to fuse a population of graphs into a single one using graph neural networks?

论文网址：Deep Graph Normalizer: A Geometric Deep Learning Approach for Estimating Connectional Brain Templates | SpringerLink

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. Tensor Representation of Multi-view Brain Networks

2.3.2. Geometric Deep Learning Layers

2.3.3. CBT Construction Layer and Subject Normalization Loss Function

2.3.4. CBT Refinement After the Training

2.4. Results and Discussion

2.4.1. Connectomic Datasets and Model Hyperparameter Setting

2.4.2. CBT Representativeness Test

2.4.3. CBT Discriminativeness Reproducibility Test

2.4.4. Discovery of Most Discriminative Brain ROIs for Each Disordered Population

2.5. Conclusion

3. 知识补充

3.1. geometric deep learning

4. Reference

1. 省流版

1.1. 心得

（1）含有可视化流程推导

（2）？受试者还能再选少一点吗

2. 论文逐段精读

2.1. Abstract

①They mentioned the average connectome, connectional brain template (CBT)

②Challenge: existing population FCs are mostly linear based rather than non-linear

③They proposed a Deep Graph Normalizer (DGN), which fuses the multi-view brain networks and captures the non-linear mode across subjects

2.2. Introduction

①Multi-modal datasets such as Human Connectome Project (HCP) and Connectome Related to Human Disease (CRHD) promote the finding of CBT (healthy)

②Briefly introduced DGN

2.3. Proposed Method

①Overall framework:

②MVBNs are $T=\{\mathbf{T}_{1}^{1},\mathbf{T}_{2}^{1},\ldots,\mathbf{T}_{i}^{v},\ldots,\mathbf{T}_{N}^{n_{v}}\}$ , where $\mathbf{T}_i^v$ means the $v$ -th view of subject $i$ .

③The same as the A) in figure, each subject $s$ can be represented as $\mathcal{T}_s\in\mathbb{R}^{n_r\times n_r\times n_v}$ ([number of ROI, number of ROI, number of view])

④Cross validation: 5 fold

2.3.1. Tensor Representation of Multi-view Brain Networks

①Edge vector $\mathbf{e}_{ij}\in\mathbb{R}^{n_{v}\times1}$ stores the view attributes

②The diagonal of the ROI matrix is set to 0

③Noded attributes matrix $\mathbf{V}^0\in\mathbb{R}^{n_\tau\times d_0}$ , they define it as identity matrix cuz there is no original node features

2.3.2. Geometric Deep Learning Layers

①Message passing operation:

$\mathbf{v}_i^l=\frac{1}{|N(i)|+1}\left(\boldsymbol{\Theta}^l.\mathbf{v}_i^{l-1}+\sum_{j\epsilon N(i)} F^l(\mathbf{e}_{ij};\mathbf{W}^l)\mathbf{v}_j^{l-1}+\mathbf{b}^l\right)$

$F^l(\mathbf{e}_{ij};\mathbf{W}^l)=\mathbf{\Theta}_{ij}$

where $\mathbf{v}_i^l$ denotes the embedding/feature of ROI $i$ at the layer $l$ , $\Theta ^l$ denotes the learnable parameter at the layer $l$ , $N(i)$ represents the neighbours of ROI $i$ , $\mathbf{b}^l\in\mathbb{R}^{d_l}$ is of course the bias term and $F^l$ maps $\mathbb{R}^{n_v}$ to $\mathbb{R}^{d_{l}\times d_{l-1}}$ （为什么会是这样的映射啊？？）, $\mathbf{W}^{l}$ denotes the weights.

2.3.3. CBT Construction Layer and Subject Normalization Loss Function

①Output obtained: $\mathbf{V}^{L} = \begin{bmatrix} \mathbf{v}_{1}^{L},\mathbf{v}_{2}^{L},...,\mathbf{v}_{n_{r}-1}^{L},\mathbf{v}_{n_{r}}^{L} \end{bmatrix}^{T}$

②They horizontally replicate $\mathbf{V}^{L} \in \mathbb{R}^{n_{r}\times d_{L}}$ to become $\mathcal{R}\in\mathbb{R}^{n_{r}\times n_{r}\times d_{L}}$

③Transposing $\mathcal{R}_{xyz}$ to $\mathcal{R}_{yxz}=\mathcal{R}^T$

④Calculating the element-wise absolute difference between $\mathcal{R}$ and $\mathcal{R}^T$ . And I guess that, I suppose there is an $\mathbf{V}^{L} \in \mathbb{R}^{n_{r}\times d_{L}}$ :

Then employ the replication operation...It actually becomes (it is $\mathcal{R}\in\mathbb{R}^{n_{r}\times n_{r}\times d_{L}}$ ):

Doing a transpose operation and we get $\mathcal{R}^T\in\mathbb{R}^{n_{r}\times n_{r}\times d_{L}}$ , it seems like...:

....a little bit different between the $\mathcal{R}\in\mathbb{R}^{n_{r}\times n_{r}\times d_{L}}$ above ahh. And finally I suppose that the element-wise absolute difference will be:

Quite a magic~Then sum them up along $z$ -axis:

We get the final CBT $\mathrm{C}\in\mathbb{R}^{n_r\times n_r}$ (I like this color scheme a lot).

⑤⭐I suppose they proposed a new loss term which called Subject Normalization Loss (SNL) Optimization. THEY DON'T WANT TO CALCULATE LOSS ON THE SAME SAMPLES ALL THE TIME, so they randomly choose a subset to measure the loss (for regularization):

$SNL_s=\sum\limits_{v=1}^{n_v}\sum\limits_{i\in S}\|\mathbf{C}_s-\mathbf{T}_i^v\|_F\times\lambda_v;\min\limits_{\mathbf{W}_1,\mathbf{b}_1...\mathbf{W}_L,\mathbf{b}_L}\frac{1}{|T|}\sum\limits_{s=1}^{|T|}SNL_s$

where $\lambda_v=\frac{\max\{\mu_j\}_{j=1}^{n_v}}{\mu_v}$ , $\mu_v$ denotes the average brain connectivity weights of view $v$

2.3.4. CBT Refinement After the Training

①They construct their CBT by select the median of all the CBTs

2.4. Results and Discussion

2.4.1. Connectomic Datasets and Model Hyperparameter Setting

（1）ABIDE

①Sample: 77, with 41 AD and 36 LMCI

②Features of subject: maximum principal curvature, the mean cortical thickness, the mean sulcal depth, and the average curvature

（2）ADNI GO

①Sample: 310 with 155 NC and 155 ASD

②Features of subject: cortical surface area, minimum principle area, and 6 cortical morphological brain networks extracted from the 4 aforementioned cortical measures

（3）Comprehensive settings

①Cross validation: 5 fold

②ROI: Desikan-Killiany atlas with 35 ROIs

③（？）Brain network: computing the pairwise absolute difference in cortical measurements between pairs of ROIs.

④Hyperparameter settings: by grid search

⑤Optimizer: Adam

⑥Learning rate: 0.0005

⑦Random samples in SNL: 10