论文网址：[2406.17086] BrainMAE: A Region-aware Self-supervised Learning Framework for Brain Signals

论文代码：https://anonymous.4open.science/r/fMRI-State-F014

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. 心得

2. 论文逐段精读

2.1. Abstract

2.2. Introduction

2.3. Approach

2.3.1. Region-aware Graph Attention

2.3.2. Brain Masked AutoEncoder

2.4. Experiments

2.4.1. Model Validation with Synthetic Data

2.4.2. fMRI Datasets

2.4.3. Pre-training Evaluation

2.4.4. Transfer Learning Evaluation

2.5. Related Work

2.6. Discussion and Conclusion

1. 心得

（1）早点发出去啊！！我要用你了！！给个名头！

（2）如果训练的人能再多一点就好了，这样看上去好像就几千个，能不能做大做强

（3）好人更一下代码！！看不到了！！拿给我用啊！！

2. 论文逐段精读

2.1. Abstract

        ①Challenges: dynamic representation of functional connectivity (FC) and noise

2.2. Introduction

        ①Modeling method: Fixed-FC and Dynamic-FC（简而言之就是前面那个是把整个BOLD序列拿来求相关矩阵，后面的就是类似滑完窗分了不同时间窗再分别求相关矩阵）

2.3. Approach

        ①Pipeline of BrainMAE:

2.3.1. Region-aware Graph Attention

        ①"There is a notable similarity in the representation properties between brain ROIs and words in language"（尊都假嘟？直接说ROI和单词很像？？也没有引用论文啊。”单独都有特征，组合在一起会变成更复杂的特征。“啊啊啊啊啊啊啊啊有点没有必要了吧）

        ②“因此，在语言建模研究的推动下，我们分配了一个可学习的d维向量，称为 ROI 嵌入”（啊.....ROI的嵌入为什么是语言来的...）全脑的ROI嵌入式

        ③Node (ROI) feature: 

        ④Node set: 

        ⑤⭐Adjacency matrix: ,  where  denotes similarity measurement and the  is asymmetry

        ⑥Attention score between nodes:

        ⑦Node feature aggregation:

2.3.2. Brain Masked AutoEncoder

        ①Time points segmentations: each window with size , and 

        ②作者设计了静态图Transient State Encoder（SG-TSE）和动态图Transient State Encoders（AG-TSE），具体区别看主图，比较明显。最终单个时间窗的输出经过线性层变成

        ③Mask: 70%

        ④Reconstruction loss: MSE:

2.4. Experiments

2.4.1. Model Validation with Synthetic Data

        ①Pre-training SG-BrainMAE on synthetic data:

2.4.2. fMRI Datasets

        ①Datasets: HCP-3T with 897 healty subjects, HCP-7T with 184, HCP-Aging with 725 and NSD with 8

2.4.3. Pre-training Evaluation

（1）Implementation Details

        ①Time points selection: 300s

        ②Segmentaion number: 20, for 15s each

        ③Mask ratio for each batch: (0, 0.8)

        ④Epoch: 1000

（2）Masked Signal Reconstruction

        ①Reconstruction visualization (A) and  performance (B):

（3）ROI Embeddings

        ①讲了上图C是他们的ROI特征聚类，D是t-sne分布展示，F和G是可视化的功能连接矩阵

（4）Representation Analysis

        ①Session separation on H and I

2.4.4. Transfer Learning Evaluation

（1）Implementation Details

        ①Cross validation: 5 fold

（2）Steady-state Variables Prediction

        ①Compared with behavior prediction model:

        ②Compared with age prediction models:
 

        ③Compared with task models:

（3）Transient Mental State Decoding

        ①Mental state classification on transient signal (about 10s)

（4）Ablation Study

        ①Difference of proposed models:

（5）Interpretation Analysis

        ①Attention score:

2.5. Related Work

        ①Lists auto encoder models and brain network analysis methods

2.6. Discussion and Conclusion

        ~