论文网址:[2307.10181] Community-Aware Transformer for Autism Prediction in fMRI Connectome (arxiv.org)
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
2.3.2. Local-global transformer encoder
2.4.1. Datasets and Experimental Settings
2.4.2. Quantitative and Qualitative Results
2.6.1. Variations on the Number of Prompts
2.6.2. Attention Scores of ASD vs. HC in Comparison between Com-BrainTF (ours) and BNT (baseline)
2.6.3. Decoded Functional Group Differences of ASD vs. HC
1. 省流版
1.1. 心得
(1)我超,开篇自闭症是lifelong疾病。搜了搜是真的啊,玉玉可以治愈但是自闭症不太行,为啥,太神奇了。我还没有见过自闭症的
(2)都学了大半年了才发现community好像是医学给的而不是自己分的...
1.2. 论文总结图
2. 论文逐段精读
2.1. Abstract
①Treating each ROI equally will overlook the social relationships between them. Thus, the authors put forward Com-BrainTF model to learn local and global presentations
②They share the parameters between different communities but provide specific token for each community
2.2. Introduction
①ASD patients perform abnormal in default mode network (DMN) and are influenced by the significant change of dorsal attention network (DAN) and DMN
②Com-BrainTF contains a hierarchical transformer to learn community embedding and a local transformer to aggregate the whole information of brain
③Sharing the local transformer parameters can avoid over-parameterization
2.3. Method
2.3.1. Overview
(1)Problem Definition
①They adopt Pearson correlation coefficients methods to obrain functional connectivity matrices
②Then divide ROIs to communities
③The learned embedding
④Next, the following pooling layer and MPLs predict the labels
(2)Overview of our Pipeline
①They provide a local transformer, a global transformer and a pooling layer in their local-global transformer architecture
②The overall framework
2.3.2. Local-global transformer encoder
①With the input FC, the learned node feature matrix can be calculated by
②In transformer encoder module,
where ,
is the number of heads
(1)Local Transformer
①They apply same local transformer for all the input, but use unique learnable tokens :
(2)Global Transformer
①The global operation is:
2.3.3. Graph Readout Layer
①They aggregate node embedding by OCRead.
②The graph level embedding is calculated by , where is a learnable assignment matrix computed by OCRead layer
③Afterwards, flattening and put it in MLP for final prediction
④Loss: CrossEntropy (CE) loss
2.4. Experiments
2.4.1. Datasets and Experimental Settings
(1)ABIDE
①Preprocessing: Configurable Pipeline for the Analysis of Connectomes (CPAC), band-pass filtering (0.01 - 0.1 Hz), no global signal regression
②Sites: 17
③Atlas: Craddock 200
④Communities: 8, cerebellum and subcortical structures (CS & SB), visual network (V), somatomotor network (SMN), DAN, ventral attention network (VAN), limbic network (L), frontoparietal network (FPN) and DMN
⑤Samples: 1009(好像是某一个的全部,似乎没有筛选) with 51.14% ASD
⑥FC: Pearson correlation
⑦Sampling strategy: stratified
(2)Experimental Settings
①Number of heads: equals to the number of communities
②Optimizer: Adam
③Learning rate: 1e-4
④Weight decay: 1e-4
⑤Train: validation: test = 0.7: 0.1: 0.2
⑥Batch size: 64
⑦Metrics: AUROC, accuracy, sensitivity, specificity
⑧Epoch: 50
2.4.2. Quantitative and Qualitative Results
(1)Comparison with Baselines (Quantitative results)
①Transformer based model: BNT
②Fixed brain network: BrainNetCNN(什么叫fixed??等于static吗?不time series吗?那不是有很多咩?看到下面有个可以学习的,这个也可以学习吧?不是卷积都是学习吗,啥意思)
③Learnable brain network: FBNETGEN
④Result demonstrability:
(2)Interpretibility of Com-BrainTF (Qualitative results)
①The averaged attention matrices of Com-BrainTF and BNT:
②Chord plot of communities with top 1% attention scores:
③ROI importance image visualized by ROI standardized attention score generated by average prompt vectors:
④Difference between important DMN and SMN in prompt vectors:
2.4.3. Ablation studies
(1)Input: node features vs. class tokens of local transformers
①They compared aodpting prompt tokens in local Transformer and combining prompt tokens and updated node features
②Comparison table:
(2)Output: Cross Entropy loss on the learned node features vs. prompt token
①They compared node features and prompt tokens as output
②Comparison table:
2.5. Conclusion
Com-BrainTF outperforms SOTA and is able to detects salient functional networks associated with ASD
2.6. Supplementary Materials
2.6.1. Variations on the Number of Prompts
Comparison/Ablation of number of input prompt tokens:
2.6.2. Attention Scores of ASD vs. HC in Comparison between Com-BrainTF (ours) and BNT (baseline)
Comparison of Com-BrainTF’s prompt vector with BNT’s attention vector:
2.6.3. Decoded Functional Group Differences of ASD vs. HC
①Biomakers in DMN, calculated by normalized attention scores:
②Biomakers in SMN, calculated by normalized attention scores:
3. 知识补充
3.1. Prompt
参考学习:通俗易懂地理解Prompt - 知乎 (zhihu.com)
3.2. Neurosynth
官网:Neurosynth
4. Reference List
Bannadabhavi A. et al. (2023) 'Community-Aware Transformer for Autism Prediction in fMRI Connectome', 26th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2023), doi: https://doi.org/10.48550/arXiv.2307.10181