[论文精读]A Federated Multigraph Integration Approach for Connectional Brain Template Learning

最新推荐文章于 2024-10-10 11:54:20 发布

夏莉莉iy

最新推荐文章于 2024-10-10 11:54:20 发布

阅读量644

点赞数 16

分类专栏：论文精读文章标签：人工智能深度学习计算机视觉笔记机器学习 embedding 神经网络

本文链接：https://blog.csdn.net/Sherlily/article/details/141901977

版权

论文精读专栏收录该内容

79 篇文章 9 订阅

订阅专栏

论文原文：A Federated Multigraph Integration Approach for Connectional Brain Template Learning | SpringerLink

论文代码：basiralab/Fed-CBT: Federated multigraph integration with application to connectional brain template estimation. (github.com)

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. A-Training c Fraction of Hospitals and Sending Models’ Weights to the Server

2.3.2. B-Global Model Update Strategy

2.4. Results and Discussion

2.4.1. Dataset

2.4.2. Hyperparameter Tuning

2.4.3. CBT Evaluation

2.4.4. Limitations and Future Directions

2.5. Conclusion

3. Reference

1. 省流版

1.1. 心得

（1）实验非常少~也不能说少吧，就是感觉那俩实验图和其他论文的实验方式差距好大，而且也没咋分析。可能是因为会议论文限制篇幅吧

（2）总觉得对自己的模型介绍的也不算清晰

2. 论文逐段精读

2.1. Abstract

①The connectional brain template (CBT) itself is a multi-view brain network

②Challenges: data privacy and sensitivity

2.2. Introduction

①Introducing CBT, models to extract CBT, data privacy, federated models and their limitations

②They proposed a federated connectional brain template learning (Fed-CBT) to sovel the problem of CBT comes from different sites. This model estimates CBT separately based on different sites

ramification n.后果；(众多复杂而又难以预料的)结果

2.3. Proposed Method

①The overall framework of Fed-CBT:

②Each epoch: each site/client randomly chooses a sub-dataset to train with shared weights, only one CBT will be send to conv layer

③Notations:

2.3.1. A-Training c Fraction of Hospitals and Sending Models’ Weights to the Server

①作者说选了n个医院，然后每轮选出c个医院参与训练，但是为什么图中参预训练的是k个医院

②Trained datasets in $k$ hospitals: $TH^t=\{TH_1^t,TH_2^t,\ldots,TH_k^t\}$ and untrained: $H^t=\{H_1^t,H_2^t,\ldots,H_{n-k}^t\}$

③Multi-view brian networks in the trained datasets: $T_{j}=\{T_{j1}^{1},\ldots,T_{ji}^{v},\ldots,T_{jn}^{n_{v}}\}$ , where the superscript denotes the view, and the second subscript denotes the subject

④Each subject can be represented as a tensor $\mathcal{T}_s^j\in\mathbb{R}^{n_r\times n_r\times n_v}$ , where $r$ is the number of ROI and $v$ is the number of view

⑤Node attributes identity matrix: $\mathbf{V}^0\in\mathbb{R}^{n_r\times d_0}$

⑥Output of the convolution layer: $\mathbf{V}_{j}^{l}=\left[\mathbf{v}_{j1}^{l},\mathbf{v}_{j2}^{l},\ldots,\mathbf{v}_{jn_{r}}^{l}\right]^{T}$

⑦To expand the x-axis of $\mathbf{V}_{j}^{l}$ and transpose it, they get $\mathcal{R}_{k}\in\mathbb{R}^{n_{r}\times n_{r}\times d_{L}}$ and $\mathcal{R}_{k}^{T}$ （三维的似乎没说是哪个维度转置？）

⑧By summing the z-axis up, the final predicted CBT will be $\mathbf{C}_j\in\mathbb{R}^{n_r\times n_r}$

⑨SNL loss:

$SNL_{js}=\sum_{v=1}^{n_v}\sum_{i\in S}\left\|\mathbf{C}_{js}-\mathbf{T}_{ji}^v\right\|_F\times\lambda_v;\min_{\mathbf{W}_1,\mathbf{b}_1\ldots\mathbf{W}_l,\mathbf{b}_l}\frac{1}{|T_j|}\sum_{s=1}^{|T_j|}SNL_{js}$

where $\lambda_{v}=\frac{\max\left\{\mu_{j}\right\}_{j=1}^{n_{v}}}{\mu_{v}}$ , the $\mu _v$ denotes the mean of brain graph connectivity weights of the view $v$

⑩All the hospitals calculate the Frobenius distance between CBT and their testing data by the latest model weights. Only trained hospital calculating the loss, uptating the model weights and backward propagating to the server. The other hospital only send the weights and loss to the server.

2.3.2. B-Global Model Update Strategy

①Loss of hospital $j$ in the end round $t$ :

$f_j(\mathcal{W}_j^t)=\frac{1}{|T_j|}\sum_{i=1}^{T_j}SNL(\mathcal{T}_{ji};\mathcal{W}_j^t)$

②Federated loss function at the round $t$ :

$f(\mathcal{W}^t)=\sum_{j=1}^Kf_j(\mathcal{W}_j^t)$

③Federated model’s weights:

$\mathcal{W}^{t+1}=\frac1N\sum_{n=1}^N\mathcal{W}_n^t$

where $\mathcal{W}_n^t$ represents the weights of all layers of the $n$ -th hospital in the $t$ -th round

④⭐To solve the problem that averaging the weights might dilute the weight/importance of more comprehensively trained model (due to the randomness of hospital selection). They adopt a temporally weighted (TW) aggregation strategy to emphasize the contribution of fully trained hospitals:

$\mathcal{W}^{t+1}=\sum_{n=1}^{N}\frac{(\frac{\exp}{2})^{-(t-ts^n)}}{\sum_{n=1}^{N}(\frac{\exp}{2})^{-(t-ts^n)}}\mathcal{W}_{n}^{t}$

where $ts^n$ denotes the latest round that the hospital $j$ trained

2.4. Results and Discussion

2.4.1. Dataset

①Dataset: ABIDE I

②Sample: 186 NC and 155 ASD

③ROI: Desikan-Killiany Atlas with 35 ROI

④Cortical hemisphere (?): “对于每个皮质半球，数据集中的每个个体都由一个连接张量表示，包括 [21,22,23] 中介绍的 6 个皮质形态学脑网络：皮质表面积、最小主面积、最大主曲率、平均皮质厚度、平均沟深和平均曲率。每个皮层视图的大脑连接权重表示它们在形态上的不同。具体来说，我们首先计算感兴趣皮层区域中特定皮层属性的平均值。接下来，我们通过计算两个区域的平均皮层属性（例如，沟深度）之间的绝对差来生成两个区域之间的连接权重。”

curvature n.曲率；弯曲；曲度

2.4.2. Hyperparameter Tuning

①Strategies: DGN w/o federation, DGN w/ federation (c=0.91, e=1), DGN w/ federation and TW (c=0.33, e=1)

②Views: ASD LH, ASD RH, NC LH and NC RH

③Validation: 3 fold cross validation

④Round number: 750 with early stop of 11 rounds

2.4.3. CBT Evaluation

①Frobenius distance (FD) between the learned CBT from the training set and all brain views in the test set:

$d_{F}(A,B)=\sqrt{\sum_{i}\sum_{j}\left|A_{ij}-B_{ij}\right|^{2}}$

②FD on different hospitals:

③ASD RH 数据集被划分为三个折叠？每个折叠进一步分为三个组？每个组分配给一个医院？

2.4.4. Limitations and Future Directions

①This method is constrained by the same amount of brain connective views

2.5. Conclusion

They want to expand their work to multi-modality

3. Reference

Bayram, H. C. & Rekik, I. (2021) 'A Federated Multigraph Integration Approach for Connectional Brain Template Learning', Multimodal Learning for Clinical Decision Support, 13050: 36-47. doi: A Federated Multigraph Integration Approach for Connectional Brain Template Learning | SpringerLink