Title
题目
Brain networks and intelligence: A graph neural network based approach toresting state fMRI data
《脑网络与智力：基于图神经网络的静息态fMRI数据研究》
01
文献速递介绍
智力是一个复杂的构念，由多种认知过程构成。 研究人员通常依赖一系列认知测试来测量不同方面的认知能力，并形成特定的智力测量指标，例如流体智力（fluid intelligence，指在新情境下进行推理和解决问题的能力）（Kyllonen and Kell, 2017），晶体智力（crystallized intelligence，指运用知识和经验解决问题的能力），以及总智力（total intelligence，总体认知能力的综合测量）。人类一直对揭示智力的神经基础以及预测个体智力差异充满兴趣。尽管传统的MRI（磁共振成像）研究主要集中在不同表型的脑结构指标上（Suresh et al., 2023; Ray et al., 2023），但近年来快速发展的研究开始利用脑功能特征来预测智力（Ferguson et al., 2017; He et al., 2020; Dubois et al., 2018）。(Vieira et al., 2022) 的综述进一步指出，功能性磁共振成像（fMRI）已成为预测智力的最常用模式，而基于静息态fMRI（rs-fMRI）的静态功能连接（FC）是最被广泛研究的预测因子。rs-fMRI通过血氧水平依赖（BOLD）信号来测量神经活动响应下的自发脑活动。功能连接（FC）被定义为通过BOLD信号时间序列计算的脑区之间的时间相关性，其能够全面反映大脑的内在组织结构（Lee et al., 2012）。研究验证了默认模式网络与额顶网络之间的功能连接对个体认知能力差异的贡献（Hearne et al., 2016）。
尽管大多数智力预测方法依赖于线性回归模型， 一些研究也采用了非线性方法，例如多项式核支持向量回归（SVR）（Wang et al., 2015）、核岭回归（He et al., 2020）以及深度神经网络（He et al., 2020; Fan et al., 2020; Li et al., 2023）。近年来，图神经网络（Graph Neural Networks, GNNs）因其在端到端图学习应用中的强大性能而受到极大关注，并快速发展。GNNs被认为是分析图结构数据的最先进深度学习方法，因为它们可以针对图中的节点和边设计神经网络，并将节点特征和边特征嵌入图的结构信息中。已有研究探讨了GNNs在社交网络、蛋白质网络以及神经生物标志物等不同领域中的有效性（Kim and Ye, 2020; Kazi et al., 2023; Nandakumar et al., 2021）。考虑到脑网络的图结构特性，通过GNN建模脑连接组已被实现。大多数脑GNN研究利用rs-fMRI的功能连接图（Ktena et al., 2018; Škoch et al., 2022; Wu et al., 2021; Ma et al., 2018），并分类研究对象的特定表型，如性别（Arslan et al., 2018; Kazi et al., 2021）或特定疾病状态（Kazi et al., 2021; Ma et al., 2018），但其在智力预测中的应用仍未被充分探索。
大多数GNN假设节点在整个图中以相同方式学习嵌入， 然而这种假设对于具有子网络特性的脑连接组而言存在问题（Parente and Colosimo, 2020）。最近，BrainGNN（Li et al., 2021）提出了一种新的GNN架构，通过在图卷积层中引入基于聚类的嵌入方法，解决了这一限制，使得不同聚类中的节点（代表不同的脑网络）可以以不同的方式学习嵌入。受BrainGNN启发，我们提出了一种新的GNN模型，称为脑感兴趣区感知图同构网络（Brain ROI-aware Graph Isomorphism Networks，BrainRGIN），用于智力预测。首先，我们使用图同构网络（GIN）（Xu et al., 2018）来增强GNN的表达能力，GIN被设计用于逼近Weisfeiler-Lehman（WL）图同构测试的能力。与BrainGNN类似，我们通过引入感兴趣区（ROI）的聚类表示解决了节点学习机制的限制。通过结合这两种架构，我们的模型能够有效捕获脑区之间的局部和全局关系。此外，我们验证了包括基于注意力的读出方法在内的各种聚合和读出函数。据我们所知，这是首个使用图神经网络结合静息态fMRI数据预测智力的研究。我们在一个大型数据集上评估了所提出模型的性能，并证明了其在预测个体智力差异方面的有效性。
Aastract
摘要
Resting-state functional magnetic resonance imaging (rsfMRI) is a powerful tool for investigating the relationship between brain function and cognitive processes as it allows for the functional organization ofthe brain to be captured without relying on a specific task or stimuli. In this paper, we present a novelmodeling architecture called BrainRGIN for predicting intelligence (fluid, crystallized and total intelligence)using graph neural networks on rsfMRI derived static functional network connectivity matrices. Extendingfrom the existing graph convolution networks, our approach incorporates a clustering-based embedding andgraph isomorphism network in the graph convolutional layer to reflect the nature of the brain sub-networkorganization and efficient network expression, in combination with TopK pooling and attention-based readoutfunctions. We evaluated our proposed architecture on a large dataset, specifically the Adolescent BrainCognitive Development Dataset, and demonstrated its effectiveness in predicting individual differences inintelligence. Our model achieved lower mean squared errors and higher correlation scores than existingrelevant graph architectures and other traditional machine learning models for all of the intelligence predictiontasks. The middle frontal gyrus exhibited a significant contribution to both fluid and crystallized intelligence,suggesting their pivotal role in these cognitive processes. Total composite scores identified a diverse set of brainregions to be relevant which underscores the complex nature of total intelligence.
静息态功能性磁共振成像（resting-state functional magnetic resonance imaging, rsfMRI）是一种强大的工具，用于研究脑功能与认知过程之间的关系，因为它能够在不依赖特定任务或刺激的情况下捕捉大脑的功能性组织。在本文中，我们提出了一种新的建模架构，称为BrainRGIN，该方法利用基于图神经网络的rsfMRI静态功能网络连接矩阵预测智力（流体智力、晶体智力和总智力）。在现有图卷积网络的基础上，我们的方法在图卷积层中结合了基于聚类的嵌入和图同构网络，反映了大脑子网络组织的特性以及高效的网络表达能力，同时结合了TopK池化和基于注意力的读出函数。
我们在一个大型数据集（特别是青少年大脑认知发展数据集，Adolescent Brain Cognitive Development Dataset）上评估了所提出的架构，并证明了其在预测个体智力差异方面的有效性。我们的模型在所有智力预测任务中相比现有相关图神经网络架构和传统机器学习模型，均实现了更低的均方误差（MSE）和更高的相关性得分。研究发现，中额回（middle frontal gyrus）对流体智力和晶体智力的预测具有显著贡献，这表明其在这些认知过程中的重要作用。而对于总智力的综合评分，识别出了一组多样化的大脑区域，这突显了总智力的复杂性。
Conclusion
结论
In this research study, a novel technique called Brain ROI-AwareGraph Isomorphism Networks, BrainRGIN , was proposed to predictintelligence using static FNC matrices derived from resting-state fMRIdata. BrainRGIN integrates the expressive power of GIN and clusteringbased GCN and also incorporates attention-based readout function,in the hope of better representing brain networks and improvingmodel prediction. Specifically, by replacing the aggregation functionof GIN with that of RGCN, the model leverages the powerful representation learning capability of GIN, while still capturing the edgestrength and edge type represented by a clustering-based embeddingmethod. Another notable aspect of the proposed architecture is the useof attention-based readout functions instead of conventional readoutmethods, which is proven to be very effective as can be seen fromTable 1. The attention mechanism assigns importance scores to eachnode, effectively capturing spatial relevance information for prediction. Using attention-based readout functions not only improved theoverall prediction of the model but also validated the theory thatdifferent brain regions contribute in a different manner to intelligencepredictionTo ensure the model is robust to influence from demographic factorssuch as age, gender, site, and socioeconomic status, we have addresseddemographic factors such as age, gender, and site (which is expected tohave a huge effect on MRI data) being controlled for the experiment.Specifically, we effectively regressed these factors from intelligencescores during model evaluation to mitigate their influence. It is noteworthy that before regressing out site from intelligence scores, the finalcorrelation scores were as high as 0.35 for fluid intelligence, 0.42 forcrystallized intelligence, and 0.41 for total composite scores with lowermean squared errors. However, removing these covariates led to a dropin overall prediction scores based on correlation and MSE metrics aspresented in Table 2. Furthermore, we did an in-depth analysis to assessthe potential impact of socioeconomic status (SES), focusing on incomeand education, on our model evaluation for crystallized intelligence.Originally, our data revealed moderate positive correlations betweenSES factors and predicted (0.28) and original crystallized intelligencescores (0.33). To further evaluate the influence of SES, we retrained themodel on crystallized intelligence, controlling for income and educationto generate new intelligence scores, keeping the model configurationconsistent with earlier training. Surprisingly, upon removing SES factors from the evaluation, the model’s performance improved slightly,as evidenced by an increased correlation of 0.331 from 0.30 on crystallized intelligence and a lower mean squared error of 256.21 from263.7. The top regions of interest were similar to the earlier modelwhere we did not regress out SES. These findings robustly demonstratethat our model evaluation remains unaffected by socioeconomic statusand other demographic factors.
在本研究中，提出了一种新的技术——脑感兴趣区感知图同构网络（Brain ROI-Aware Graph Isomorphism Networks, BrainRGIN），用于通过静息态fMRI数据提取的静态功能网络连接（FNC）矩阵预测智力。BrainRGIN 将GIN（图同构网络）的强大表达能力与基于聚类的GCN（图卷积网络）相结合，同时引入了基于注意力的读出函数，以更好地表征脑网络并提高模型预测能力。具体而言，通过将GIN的聚合函数替换为RGCN的聚合函数，该模型不仅利用了GIN的强大表示学习能力，同时通过基于聚类的嵌入方法捕获了边强度和边类型的特性。该架构的另一个显著特点是使用基于注意力的读出函数代替传统的读出方法。从表1可以看出，基于注意力的读出函数非常有效。注意力机制为每个节点分配重要性分数，有效地捕捉了空间相关性信息以进行预测。使用基于注意力的读出函数不仅改善了模型的整体预测性能，还验证了不同脑区在智力预测中具有不同贡献的理论。为了确保模型对人口学因素（如年龄、性别、采集站点和社会经济地位）的影响具有鲁棒性，实验中已对这些因素进行控制。特别是在模型评估过程中，我们对智力分数中这些因素进行了回归处理，以减少其影响。值得注意的是，在对智力分数回归掉采集站点的影响之前，流体智力的相关性得分高达0.35，晶体智力为0.42，总复合得分为0.41，且均方误差较低。然而，去除这些协变量后，相关性和MSE指标显示整体预测分数有所下降（见表2）。
此外，我们对社会经济地位（SES）的潜在影响进行了深入分析，重点关注收入和教育对模型在晶体智力预测中的评估影响。最初的数据表明，SES因素与预测的晶体智力分数（相关性为0.28）和原始晶体智力分数（相关性为0.33）之间存在中等正相关性。为了进一步评估SES的影响，我们控制了收入和教育因素，重新训练了晶体智力模型，并生成了新的智力分数，同时保持模型配置与之前训练一致。令人惊讶的是，去除SES因素后，模型性能略有提升，晶体智力的相关性从0.30提高到0.331，均方误差从263.7下降到256.21。与未回归SES因素的模型相比，主要感兴趣区域保持相似。
这些发现强有力地表明，我们的模型评估结果不受社会经济地位及其他人口学因素的影响。
Results
结果
The experimental results are summarized in Tables 1 and 2. InTable 1, it is evident that the BrainRGIN architecture demonstratedpromising performance in predicting fluid intelligence, achieving amean squared error (MSE) of 263 and a correlation coefficient of 0.23when employing RGIN convolution combined with the SERO readoutmethod. Moreover, it yielded the best results in predicting crystallizedintelligence, with an MSE of 263.7 and a correlation of 0.30 using thesame RGIN convolution and SERO readout method. Notably, for totalcomposite scores, the GARO attention-based readout function in conjunction with the RGIN graph model attained the highest performance,achieving an MSE of 261 and a correlation of 0.31.Furthermore, we observed stable and comparable performancesfrom baseline models such as BrainNetCNN (Kawahara et al., 2017),FBNetGen (Kan et al., 2022a), and GT (Dwivedi and Bresson, 2020),indicating the reliability of these models. However, the Brain NetworkTransformer (BNT) (Kan et al., 2022b) exhibited robust performancewith higher correlation scores and lower mean squared errors. AlthoughBrainRGIN occasionally reported lower mean squared errors comparedto BNT, the performance appeared to be comparable. Additionally,BrainRGIN* surpassed all traditional baseline models and exhibitedsuperior performance compared to Support Vector Regression (SVR),Linear Regression (LR), and Ridge Regression. It consistently achievedlower MSE values and higher correlation coefficients across all metrics,namely fluid intelligence, crystallized intelligence, and total compositescores. These results underscore the effectiveness of RGIN aggregation and attention-based readout methods over the BrainGNN modeland other baseline models in predicting intelligence scores. Moreover,attention-based readout methods were found to outperform other readout techniques, significantly contributing to the architecture’s successin predicting intelligence scores.These findings underscore the importance of carefully selecting appropriate graph neural network components for predicting intelligencescores from rsfMRI data and provide a valuable foundation for futureresearch in this area.
实验结果总结如表1和表2所示。在表1中，可以明显看出，BrainRGIN 架构在智力预测中表现出良好的性能。当采用RGIN卷积和SERO读出方法时，其在流体智力预测中实现了均方误差（MSE）为263，相关系数为0.23。此外，在晶体智力预测中也取得了最佳结果，MSE为263.7，相关系数为0.30，同样使用RGIN卷积和SERO读出方法。值得注意的是，在总智力复合得分预测中，结合RGIN图模型的GARO注意力读出函数达到了最高性能，MSE为261，相关系数为0.31。
此外，我们观察到基线模型如BrainNetCNN（Kawahara et al., 2017）、FBNetGen（Kan et al., 2022a）和GT（Dwivedi and Bresson, 2020）表现出稳定且可比的性能，表明这些模型的可靠性。然而，Brain Network Transformer（BNT）（Kan et al., 2022b）表现出较强的性能，具有更高的相关性分数和更低的均方误差。尽管BrainRGIN 在某些情况下报告的均方误差低于BNT，但整体性能相当。此外，BrainRGIN 超越了所有传统的基线模型，并在支持向量回归（SVR）、线性回归（LR）和岭回归的性能上也表现出优越性。在流体智力、晶体智力和总复合得分的所有指标中，BrainRGIN 一致实现了更低的MSE值和更高的相关系数。这些结果表明，RGIN聚合方法和基于注意力的读出方法在智力分数预测中的有效性优于BrainGNN模型和其他基线模型。此外，基于注意力的读出方法优于其他读出技术，这对架构在智力分数预测中的成功起到了显著的作用。
这些发现强调了在从静息态fMRI数据中预测智力分数时，精心选择合适的图神经网络组件的重要性，并为该领域的未来研究提供了有价值的基础。
Figure
图
Fig. 1. Overall architecture of BrainRGIN. The static FNC matrix is extracted from a resting state fMRI time series data. Three blocks of BrainRGIN are used with attention-basedreadout functions followed by a fully connected layer for prediction.
图1. BrainRGIN 的整体架构。静态功能网络连接（FNC）矩阵从静息态fMRI时间序列数据中提取而来。模型中包含三个BrainRGIN模块，每个模块采用基于注意力的读出函数，最终通过一个全连接层进行预测。
Fig. 2. Regions significant in fluid and crystallized intelligence prediction.
图2. 在流体智力和晶体智力预测中具有显著意义的脑区。
Fig. 3. Significant regions expressed as connectivity networks for total composite scores.
图3. 表现为总复合得分连接网络的显著脑区。
Table
表
Table 1Comparison of different BrainRGIN architectures for intelligence prediction.
表1 不同 BrainRGIN 架构在智力预测任务中的比较。
Table 2Comparison of BrainRGIN with baseline models on ABCD dataset
表2 BrainRGIN 与基线模型在ABCD数据集上的比较。
Table 3Effect of edges threshold selection in model prediction.
表3 边阈值选择对模型预测的影响。
Table 4Comparison of BrainRGIN with baseline models on HCP dataset.
表4 BrainRGIN 与基线模型在HCP数据集上的比较。
Table 5Evaluation of BrainRGIN components and their performance on total composite scores.
表5 BrainRGIN 组件及其在总复合得分上的性能评估。
