[论文精读]Classification of schizophrenia patients using a graph convolutional network: A combined fun-

①Acquisition: "using an echo-planar imaging (EPI) sequence (matrix = 64 × 64 × 36; spatial resolution = 3.4 × 3.4 × 4 mm3; echo time (TE) = 30 ms; acquisition time = 2000 ms; flip angle = 90°; field of view = 211 × 211 mm2)" by Philips 3.0T MR system

②State of subjects: "lie calmly in the instrument, breathe smoothly, and keep their eyes closed but remain conscious"

③Preprocessing softwares: SPM 8 abd DPABI

④一段医学影像处理的话，我没办法paraphrase了，而且我处理也没学那么好："The ALFF method was used to measure regional spontaneous neuronal activity. The time series of each voxel was transformed to the frequency domain with a fast Fourier transform (FFT), and the power spectrum was obtained. ALFF was calculated as the averaged square root across 0.01–0.08 Hz and then divided by the global mean ALFF value for each subject for standardization purposes. The ALFF value of the ROI was represented by the average ALFF value of all voxels in the ROI. ReHo measures the functional synchronization of a voxel with its nearest neighbors. In this paper, the ReHo value of a voxel was represented by Kendall’s coefficient of concordance (KCC) of the voxel’s time series with its 26 nearest neighbors. The ReHo value of each voxel was then divided by the global mean ReHo, and the ReHo value of the ROI was defined in the same way as ALFF analyses."

2.3.3. Construction of a functional brain graph

①Brain graph construction:

②Nodes: ROIs

③Edges: functional connectivities (FC) calculated by absolute Pearson correlation coefficient（emm绝对值这样好吗，感觉正负相关不能一概而论诶）

④Atlas: AAL 90

⑤The authors selected edges with significant proportions of weight to ensure the topological structure（你怎么不说这个阈值是多少呢？？）

⑥Functional characteristics: Based on my guess, thier feature matrix can be $H \in \mathbb{R}^{N\times M}$ , where $N$ denotes the number of ROIs and $M$ denotes the number of features. What is more, they choose ALFF, ReHo and DC as node features（就像上图蓝色，应该是每个节点三个特征）. DC in this paper is defined as the sum of all the FC of one ROI

⑦Node property characteristics: However, in comparison experiments, ALFF and ReHo are replaced by nodal efficiency (NE) and betweenness centrality (BC), which are network property features（什么对比实验？消融实验吗？奇奇怪怪的用词）

spurious adj.虚假的;伪造的;谬误的;建立在错误的观念(或思想方法)之上的

2.3.4. Proposed methods

①看烂了的function of GCN layer:

$H^{(l+1)}=\text{ReLu}\bigg(\widetilde{D}^{-\frac{1}{2}}\widetilde{A}\widetilde{D}^{-\frac{1}{2}}H^{(l)}W^{(l)}\bigg)$

②To achieve pooling operator, they weed out some nodes after convolution by top-k pooling layer（后面紧接一个k=ratio*N我真是。我。难绷）

③Each pooling layer follows a READOUT layer, which combines global max-pooling with global mean-pooling

④Overall framework:

2.4. Experiment and results analysis

2.4.1. Experimental settings

①They randomly divide training set and test set to 285 and 60, which is 4.75:1（哈哈哈哈哈哈哈哈哈哈作者非常诚实地说自己进行了两百次随即拆分）

②Parameter setting:

③Metrics: ACC, AUC, SEN, SPE, PR, F1-score

④Maximum epoch: 100k, but early stops when there is no vibration on loss in 200 epochs

dispersion n.分散;散布;散开 oscillate vi.振荡;振动;摆动;(情感或行为)摇摆，变化;波动

2.4.2. Construction of a sparse binary network

①They test each dropout ratio of edges in functional characteristics:

when keeps 7% edges, the performance achieves the best

②Dropout ratio of edges in node property characteristics:

when keeps 15% edges, the performance achieves the best

2.4.3. Results of the same model initialization

①The ablation study of different 200 types of training/testing set:

②Average performance over 200 model on the same model initialization:

啊这么高啊

2.4.4. Results of random model initialization

①什么是模型初始化？为什么不细🔒一下。是参数吗，咋随机啥范围也不提啊？还是说工具随机啊？

②Average performance over 200 model on the random model initialization:

③"Statistical analysis of the difference values of accuracy and AUC between the random initialization and the same initialization":

④"The curves of the difference value of accuracy and AUC":

2.4.5. Comparisons with other methods

①Comparison table:

unimodal adj.（曲线）单峰的

2.4.6. Limitations

①Limit number of samples

②Different atlas arenot attempted

③Multi-modality is not attempted

2.5. Conclusions

They achieve excellent performance without model name...

3. 知识补充

3.1. ALFF/ReHo

ALFF（Amplitude of Low-Frequency Fluctuation）和ReHo（Regional Homogeneity）都是用于分析静息态功能磁共振成像（fMRI）数据的指标，它们能够反映大脑在静息状态下的功能活动情况。

（1）ALFF是低频振幅，表示功能活动的强度。在fMRI数据中，每个体素的BOLD信号都会随时间发生变化，ALFF就是计算这些信号在低频段（通常认为是0.01~0.1Hz）的振幅大小。这个指标可以反映大脑各个区域的自发活动水平。

ALFF得到的结果不是一个向量，而是一个表示大脑不同区域功能活动强度的数值或图像。具体来说，ALFF通过分析脑电信号或功能磁共振成像（fMRI）信号的低频振荡幅度，来研究脑区的功能活动。在预处理和计算过程中，通过对数据的处理和分析，最终可以生成一个三维ALFF图像（那看起来在文中的功能脑图构建图中ALFF是被展平了的），这个图像可以显示大脑不同区域的ALFF值。此外，还可以生成统计图来比较不同组或条件之间的差异，或者使用区域感兴趣区（ROI）分析来提取特定脑区的ALFF值。

（2）ReHo是局部一致性，它反映了邻近体素间BOLD信号变化的同步性。在一个特定的体素周围，如果邻近体素的BOLD信号变化高度一致，那么该体素的ReHo值就会较高。这个指标可以用于衡量大脑局部功能连接的强度。

4. Reference List

Chen, X. et al. (2023) 'Classification of schizophrenia patients using a graph convolutional network: A combined functional MRI and connectomics analysis', Biomedical Signal Processing and Control, 80 (1), 104293. doi: Redirecting

标签

#笔记 #人工智能 #机器学习 #深度学习 #学习