论文全名:Hierarchical Graph Convolutional Network Built by Multiscale Atlases for Brain Disorder Diagnosis Using Functional Connectivity
论文网址:基于多尺度图谱构建的分层图卷积网络,用于使用功能连接进行脑部疾病诊断 |IEEE期刊和杂志 |IEEE Xplore(IEEE Xplore)
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
目录
2.3.1. Subjects and rs-fMRI Data Preprocessing
2.3.2. Multiscale Atlases for Brain Parcellation and FCN Construction
2.3.3. Multiscale-Atlases-Based Hierarchical Graph Convolutional Network (MAHGCN)
2.3.4. Model Explainability Analysis
2.5.2. Comparison With State-of-the-Art Methods
2.5.4. Explainability Analysis
1. 省流版
1.1. 心得
(1)外行表示不同Atlas真有那么大区别吗?得看按啥标准分的?
(2)⭐用统计学方法测量差异性从而选出主要参评数据集
(3)⭐对于不同的Atlas有待考究
(4)⭐咦,和SOTA比较不需要同样的样本数吗
(5)后面用柱状图表示的实验还把精度写在了正文里...其实可以做个表清楚一点
(6)好的好的看完代码对你意见没那么大了
1.2. 论文总结图
2. 论文逐段精读
2.1. Abstract
①Exsting works construct graph with single scale. Ergo, the authors introduced multi-scale graph
②They proposed Atlas-guided Pooling (AP) strategy
③They put forward a multiscale-atlases-based hierarchical graph convolutional network (MAHGCN)
2.2. Introduction
①Methods of constructing functional connectivity (FC): Pearson correlation, seed based correlation, and independent component analyses (ICA)
②”多尺度大脑网络不是单尺度结构的简单复制,而是大脑系统本质上的不同方面。研究发现,不同尺度的健康脑连接体具有不同的拓扑属性,如度分布[14]、中心性[11]、小世界性[15]等“(持怀疑态度)
③Briefly introducing some pool strategies, such as DIFFPOOL, Graph-U-Net, gPOOL, and SAGPOOL
2.3. Materials and Methods
2.3.1. Subjects and rs-fMRI Data Preprocessing
①Datasets: ADNI (ADNI-2 and ADNI-GO), OASIS-3, ABIDE I
②⭐Applying Chi-square test and two-sample t-test to verify there is no significant impact on sex and age
③⭐However, there are significant differences in MMSE between HC and AD. Hence the authors adopt ADNI as the main dataset
④ADNI: 531 samples with 364 HC, 117 eMCI, and 50 AD
⑤Scanning parameters of ADNI: TR = 3 s, TE = 30 ms, flip angle = 80◦ , resolution = 3.3 × 3.3 × 3.3 mm3 , 48 axial slices, and 420 s in duration (140 volumes)
⑥OASIS: 250 samples with 207 HC and 43 AD
⑦Scanning parameters of OASIS: TR = 2.2 s, TE = 27 ms, resolution = 4 × 4 × 4 mm3 , 36 axial slices, and 372 s in duration (169 volumes)
⑧ABIDE: 1011 samples with 512 HC and 499 ASD (excluding time<300s)
⑨Scanning parameters of ABIDE: TR = 1.5–3 s, TE = 15–33 ms, in-plane resolution 3 × 3–3.438 × 3.438 mm2 , slice thickness = 3–4.5 mm, 28–40 axial slices, and 304–486 s in duration (120–300 volumes)
⑩Dataset table:
⑪Introducing how they process the data
2.3.2. Multiscale Atlases for Brain Parcellation and FCN Construction
①Atlas choosing: 100, 200, 300, 400, 500
2.3.3. Multiscale-Atlases-Based Hierarchical Graph Convolutional Network (MAHGCN)
(1)Graph Convolutional Network
①Laplacian matrix:
②GCN:
where is feature matrix, denotes graph convolutional kernel, denotes learnable parameter and represents Fourier transform
③Simplifying this operation to:
where denotes ReLU(其实①②③都没有什么写上去的必要吧,也可以直接写③)
(2)Atlas-Guided Pooling
①Re mapping the matrix by voxel overlapping ratio with threshold Th(体素的重叠通过Schaefer模板算):
这样?
②The computing method of AP:
那这就是(P×R)*(R*chin)=P×chin(但是似乎不能固定留下来几个点,还是得看Th)
③⭐They adopt across scales sum pooling(就是把500*1,400*1,...,100*1的二维张量concatenate起来)
(3)Stacks of GCNs and APs for Hierarchical Architecture
①⭐输入节点特征矩阵是500×500 one hot编码,意思是chin也是500,但是chin按照上面那个AP来说不会变啊。作者也没说是GCN变得,说是AP转换的。啊????
②Overall framework:
2.3.4. Model Explainability Analysis
①⭐To evaluate the importance of each ROI, they introduce gradient-guided class activation map (GradCAM) to calculate Grad-CAM on different scales.
2.4. Experiments
2.4.1. Experimental Settings
(1)Implementation
①Loss: weighted cross entropy loss(为什么是为了解决positive和negative的sample size???). And the weights are set as the inverse of the sample ratio in the training set
②Epoch: 100
③Learning rate: 0.001
④Batch size: 30
⑤Optimizer: Adam
⑥Weight decay: 0.1 for ADNI and 0.01 for others
(2)Validation Scheme and Evaluation Metrics
①Running times: 5
②Training/testing set: 0.8/0.2 randomly
③Metrics: ACC, SEN, SPE, AUC
(3)Statistical Analysis
①Applying one-sided paired t-tests and Wilcoxon signed-rank test to verify the effectiveness of their model
2.4.2. Comparison Methods
①They compare their model with baseline (GCN, ML, RNN), pooling strategies, and multi-scale fusion methods
2.5. Results
2.5.1. Comparison Study
①Comparison table:
②The performance of multi scale model>single
③⭐“所有的最佳性能都是由基于多尺度地图集的方法获得的,而不是数据驱动的池化方法,这表明基于生物先验的方法比数据驱动的方法在正确捕获多尺度大脑特征方面具有优势。”(这句话挺精炼明了了,我就懒得再paraphrase了)
④⭐Data may influenced by noise and atlases bring hierarchical prior
⑤There is generalizability of their model
2.5.2. Comparison With State-of-the-Art Methods
①Comparison table:
2.5.3. Ablation Study
(1) Effects of the Range of Integrated Scales
①T-test and (∗ ) and Wilcoxon signed-rank test (◦ ) present in (a) HC-versus-eMCI and (b) HC-versus-AD classification on different scales:
②⭐However, there is ceiling in growth of scales
(2)Different Configurations of AP
①Comparison table with different pooling strategies:
(3)Effects of Skip Connections
①Comparison table with skip connection strategies:
(4)Effects of Thresholds for Atlas Mapping Matrix
①⭐They define Th=0, which means they contains all the overlapping
②Performance under different Th:
(5)Effects of Preprocessing Pipeline
①Displayed performances under two different preprocessing pipelines
2.5.4. Explainability Analysis
①Activation value distribution:
2.6. Conclusion
①They proposed a novel multi-scale method
3. 知识补充
3.1. Grad-CAM
论文解读:[论文总结] Grad-CAM - 知乎 (zhihu.com)
4. Reference List
Liu, M. et al. (2023) 'Hierarchical Graph Convolutional Network Built by Multiscale Atlases for Brain Disorder Diagnosis Using Functional Connectivity', IEEE Trans Neural Netw Learn Syst. doi: http://10.1109/TNNLS.2023.3282961