[论文精读]FBNETGEN: Task-aware GNN-based fMRI Analysis via Functional Brain Network Generation

论文网址：https://arxiv.org/abs/2205.12465

论文代码：https://github.com/Wayfear/FBNETGEN

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

1. 省流版

1.1. 心得

1.2. 论文总结图

2. 论文逐段精读

2.1. Abstract

2.2. Introduction

2.3. FBNetGen

2.3.1. Overview

2.3.2. Time-series Encoder

2.3.3. Graph Generator

2.3.4. Graph Predictor

2.3.5. End-to-end Training

2.4. Experiments

2.4.1. Experimental Settings

2.4.2. Performance Comparison and Ablation Studies

2.4.3. Interpretability Analysis

2.5. Conclusion 

2.6. Appendix A. Background and Related Work

2.6.1. fMRI-based Brain Network Analysis

2.6.2. Graph Neural Networks

2.7. Appendix B. Experimental Settings

2.7.1. Dataset

2.7.2. Metrics

2.7.3. Implementation details

2.7.4. Computation complexity

2.8. Appendix C. The Design of Two Encoders

2.9. Appendix D. Influence of Hyper-parameters

2.10. Appendix E. Influence of Pooling Strategies

2.11. Appendix F. Training Curves of FBNetGen Variants

2.12. Appendix G. Abbreviations of Neural Systems

2.13. Appendix H. Difference Score T of Functional Modules

2.13.1. The Definition of Difference Score T

2.13.2. Difference Score T of Functional Modules on Learnable Graph and Pearson Graph

3. Reference List

---

1. 省流版

1.1. 心得

（1）⭐“增加窗口大小和嵌入大小并不一定会提高FBNetGen的整体性能，说明FBNetGen的性能是稳定的”。哈哈哈哈哈哈

（2）⭐在2.13.2. 中，感觉皮尔逊和他们的方式都没有完美契合医学。在文中作者也是说别人的其实没太真正契合。真的吗，好奇怪，我怎么看别人的论文夸夸夸都说自己的显著区间和医学一样。

1.2. 论文总结图

​​​​​​​

2. 论文逐段精读

2.1. Abstract

        ①There are noise, unknown of prediction task and incompatible with GNN in traditional brain network（传统的功能性脑网络是啥？和GNN不兼容？不知道下游预测任务？有噪音？）

        ②They provide feature extraction of ROI, brain network generation and prediction by brain network

2.2. Introduction

        ①Researchers only know the difference between sexes, but can not explain the reason

        ②Brain network only based on correlation matrices will lack time series. Moreover, most of the GNN is incompatible with brain network with positive and negative edges（emm...这个不兼容会不会有点夸张了？虽然说确实是个小问题）

        ③They put forward an end-to-end differentiable pipeline from BOLD signal series to clinical predictions, which includes a) a time-series encoder for reducing the dimension and noise of original data, b) a graph generator from encoded features, c) a GNN predictor based on generated network

        ④FBNetGen performs good in sex prediction in  Adolescent Brain Cognitive Development (ABCD) and PNC

2.3. FBNetGen

2.3.1. Overview

        ①The overall framework of FBNetGen:

        ②The input of FBNetGen is , where  is the number of samples,  denotes the number of ROIs and  denotes time series.  is an individual.

        ③The final output is the prediction label , where  is the class set

        ④Apply functional brain network 

2.3.2. Time-series Encoder

        ①As traditional encoding methods, ICA and PCA ignore the time series

        ②Pearson correlation contains negtive weight in edges

        ③They apply 1D-CNN and bi-GRU in encoding and note the expression of encoder as: 

2.3.3. Graph Generator

        ①⭐Different from original Pearson correlation, they design ,  to eliminate negative edges

        ②The over lengthy of gradient feedback in deep learning will hinder the performance

        ③They designed 3 regularizers, group intra loss, group inter loss and sparsity loss.

        ④Group intra loss: ⭐designed for minimize the differences between the same group. It integrates samples with the same label as a group, namely  denotes the class,  denotes the samples with the same label . Then mean  and variance  can be calculated by:

and then the group intra loss is:

        ⑤Group inter loss: ⭐designed for maxmize the differences between different groups. The loss can be calculated by:

        ⑥Sparsity loss: there are some large values calculated by the two losses above, then they proposed sparsity loss to reduce the deviation:

skewed  adj. 歪斜的；曲解的    substantiate  vt. 证实；使实体化

2.3.4. Graph Predictor

        ①The node feature  is the time series

        ②The uppdate function is  and 

        ③The final classification is 

2.3.5. End-to-end Training

        ①The overall loss is , where  denotes the supervised cross-entropy loss, and , ,  are hyper-parameters

2.4. Experiments

2.4.1. Experimental Settings

（1）Philadelphia Neuroimaging Cohort (PNC) dataset

        ①Samples: 503 subjects with 289 (57.46%) female

        ②Atlas: 264 ROIs (Power et al., 2011)

（2）Adolescent Brain Cognitive Development Study (ABCD) dataset

        ①Samples: 7901 with 3961 (50.1%)

        ②Atlas: HCP 360 ROIs

（3）Metrics

        ①AUROC

        ②Accuracy

        ③The results all take from the mean value of 5 runs

2.4.2. Performance Comparison and Ablation Studies

        ①They compared their model with four types of models: time-series encoding only, traditional models, another generators and another deep models.

        ②Table of comparison:

        ③Ablation study of each loss:

2.4.3. Interpretability Analysis

        ①Heatmap of their matrices and matrices of Pearson correlation:

        ②The significance level in their functional matrices is close to biological research

        ③They adopt t-test to test the significance level and verify the results of thier model matches the biological research more

2.5. Conclusion 

        They proposed a taskk-aware GNN-based framework, FBNetGen

2.6. Appendix A. Background and Related Work

2.6.1. fMRI-based Brain Network Analysis

        ①Introduce functional connectivity (FC) in neuroscience and prove its effectiveness and importance

2.6.2. Graph Neural Networks

        ①Briefly introduce various GNN

        ②Most of GNNs require distinct graph structures and node features, which are not suitable in brain network

2.7. Appendix B. Experimental Settings

2.7.1. Dataset

（1）PNC

        ①10 functional modules are divided in 264 ROIs

        ②Preprocessing: despiking, slice timing correction, motion correction, registration to MNI 2mm standard space, normalization to percent signal change, removal of linear trend, regressing out CSF, WM, and 6 movement parameters, bandpass filtering (0.009{0.08), and spatial smoothing with a 6mm FWHM Gaussian kernel

        ③Time points: 120

（2）ABCD

        ①Time points: choose the samples which time points more than 512 and only adopt the first 512 steps (selecting)

        ②Functional module: 6

2.7.2. Metrics

        ①Due to the balanced across classes characters of PNC and ABCD, AUROC is a fair metric for them

2.7.3. Implementation details

        ①The number of conv layers in 1D-CNN: 3

        ②The number of conv layers in GRU: 4

        ③Adopting grid search in hyper-parameters

        ④Parameters in 1D-CNN:

        ⑤Hyper-parameters in loss: 

        ⑥Training set: 70%, validation set: 10%, test set: 20%

        ⑦Learning rate: 

        ⑧Weight decay: 

        ⑨Batch size: 16

        ⑩Epochs: 500

2.7.4. Computation complexity

        ①Computation complexity of feature encoder: 

        ②Computation complexity of graph generator: 

        ③Computation complexity of graph predictor: 

where  denotes the layer of feature encoder,  denotes the number of ROIs,  denotes the length of time-series

        ④The overall cc: 

2.8. Appendix C. The Design of Two Encoders

        ①The specific process of encoder in 1D-CNN:

where , the kernal size in  is the window size 

        ②The conv layers in bi-GRU:

where  represents splitting input sequence  into  segments of length . Then followed by a MLP layer: 

2.9. Appendix D. Influence of Hyper-parameters

        ①They discuss two hyper-parameters with the greatest influences, the window length  and embedding size 

        ②Figures of hyper parameters varying:

2.10. Appendix E. Influence of Pooling Strategies

        ①Ablation of pooling strategies:

2.11. Appendix F. Training Curves of FBNetGen Variants

        ②Ablation of loss function in training curves:

2.12. Appendix G. Abbreviations of Neural Systems

        ①On PNC, SH (Somatomotor Hand), SM (Somatomotor Mouth), Sub (Subcortical), Vis (Visual), Aud (Auditory), CO (Cingulo-opercular), Sal (Salience), DMN (Default mode), FP
(Fronto-parietal), VA (Ventral attention), DA (Dorsal attention), MR (Memory retrieval),

        ②On ABCD, SM(Somatomotor), DMN (Default mode), VS (Ventral salience), CE (Central
executive), DS (Dorsal salience), Vis (Visual). In ABCD, SM (Somatomotor Mouth), DMN
(Default mode), VS (Ventral salience), CE (Central executive), DS (Dorsal salience), Vis
(Visual).

2.13. Appendix H. Difference Score T of Functional Modules

2.13.1. The Definition of Difference Score T

        ①"The difference score  of each predefined functional module  is calculated as:

where  denotes the set containing all indexes of nodes belonging to the module " （呃呃怎么催生出一种割裂感，这是什么玩意儿，然后这个字符1又是什么为啥没有解释）

2.13.2. Difference Score T of Functional Modules on Learnable Graph and Pearson Graph

        ①The significance level of each functional module in PNC, while the bold words are significant modules in neurobiological findings:

        ②The significance level of each functional module in ABCD:

3. Reference List

Kan X. et al. (2022) 'FBNETGEN: Task-aware GNN-based fMRI Analysis via Functional Brain Network Generation', MIDL. doi: https://doi.org/10.48550/arXiv.2205.12465

Power J. et al. (2011) 'Functional Network Organization of the Human Brain', Neuron, 72, pp. 665-678.