磁共振成像（Magnetic Resonance Imaging, MRI）技术的快速发展，为研究活体人脑结构和功能提供了重要技术手段。随着神经影像大数据的出现，结合神经影像和机器学习的个体脑龄预测为实现老年脑健康智能评估提供了新颖的思路，然而目前的脑龄预测研究主要基于局部脑影像特征，忽略了不同脑区之间结构和功能连接的信息。通过神经影像人脑连接组学方法可从全脑系统水平对宏观尺度下大脑结构和功能网络进行建模，进一步提取对年龄敏感的脑连接特征进行个体脑龄预测。特别图卷积神经网络（Graph Convolutional Network, GCN）作为一种基于图结构的深度学习算法，可充分利用脑区连接信息，自主学习脑网络特征。因此，本研究通过利用多模态脑网络数据，结合GCN算法，构建个体脑龄预测模型，并在传统GCN模型中加入注意力机制，对用于聚合节点信息的邻接矩阵进行参数化，构建了图注意神经网络（Graph Attention Network, GAT）；进一步，我们基于多模态脑网络数据融合，构建多通道脑龄预测模型，并系统比较单一模态脑网络数据以及多模态脑网络融合对脑龄预测精度的影响；此外，对比了卷积神经网络（Convolutional neural network, CNN）、GCN和GAT三种不同深度学习算法的预测性能。

通过对比多种预测模型，实验结果发现基于白质结构脑网络的预测精度要高于功能脑网络，且多模态脑网络融合的预测精度高于任意单模态数据，表明了白质结构脑网络包含了更多对年龄敏感的拓扑特征，而且不同模态影像可提供有利于脑龄预测的特异性互补信息。对比不同的深度学习算法，我们发现GAT模型（r=0.756,
MAE=5.104）的脑龄预测精度最高，GCN模型的预测结果（r=0.743, MAE=5.164）优于CNN模型（r=0.686, MAE=5.392），进一步，通过将预测年龄与实际年龄的差异（The predicted age difference, PAD）与中老年个体认知评分进行相关分析，发现PAD与个体被试的执行功能（r=-0.273,
p=0.038）和推理能力（r=-0.366,
p=0.030）具有显著相关性，且在75岁以上高龄被试中相关性更加显著。

最后，对上述研究中使用的基于MRI数据的多模态脑网络建模方法，局部和全局脑网络特征提取和基于机器学习的分类预测算法进行整合，开发了一款基于MATLAB的“一站式”脑网络建模和个体预测软件（BCMIP），对不同模态的MRI数据处理提供了统一的输入输出数据格式，能够自动构建多模态脑网络，从而增进对多模态影像信息的应用；联合不同模态的影像信息，基于支持向量机等一系列高效、前沿的机器学习算法，对特定问题进行拆分训练、构建预测模型。该软件的开发方便了神经影像人脑连接组的计算建模和大数据挖掘，今后可应用脑健康智能评估，脑疾病智能诊断等多种问题的分类预测研究中。

The development
of Magnetic Resonance Imaging (MRI) technology provides an important technical
means for studying the structure and function of the living human brain. With
the emergence of neuroimaging big data, individual brain age prediction
combined with neuroimaging and machine learning provides a novel idea for
realizing the intelligent assessment of elderly brain health. Combined with
neuroimaging human brain connectomics method, individual brain age can be
predicted based on the whole-brain level considering human brain structure and
functional connectivity information. In particular, Graph Convolutional Network
(GCN) can make full use of brain area connection information and learn brain
network features autonomously. Therefore, this study constructs an individual
brain age prediction model through multimodal brain network data,
systematically compared the effects of single-modality brain network data and
multi-modality brain network fusion on the accuracy of brain age prediction,
and the performance of three different deep learning algorithms, Convolutional
Neural Network (CNN), GCN and Graph Attention Network(GAT).

The experimental
results show that the prediction accuracy of the brain network based on the
white matter structure is higher than that of the functional brain network, and
the prediction accuracy of the multimodal brain network fusion is higher than
that of any single-modal data. Comparing different deep learning algorithms, we
found that the GAT model (r=0.756, MAE=5.104) had the highest prediction
accuracy for brain age, and the GCN model (r=0.743, MAE=5.164) was better than
the CNN model (r=0.686, MAE=5.392). By correlating the difference between
predicted age and actual age (The predicted age difference, PAD) with the
cognitive scores of middle-aged and elderly individuals, we found that PAD was
closely related to the executive function of individual subjects (r=-0.273,
p=0.038) and reasoning ability (r=-0.366, p=0.030), and the correlation was
more significant in subjects over 75 years old.