Abstract:

In view of the low accuracy of functional Magnetic Resonance Imaging （fMRI） brain age prediction and the lack of research on the combination of this problem and deep learning， an fMRI brain age prediction model with Lightweight Multi-scale Convolutional Network （LMCN） was proposed. Firstly， the Pearson correlation coefficient （R） of the Region Of Interest （ROI） in fMRI was calculated to obtain the Functional Connectivity （FC） matrix of the ROI as input. Secondly， the number of FC channels was increased to ensure the number of features and the size of the feature map was reduced simultaneously. At the same time， the multi-scale dilated convolution module RFB （Receptive Field Block） with the characteristics of human visual attention was used to extract age features. Finally， the predicted brain age was output by the fully connected layer， and the ablation prediction results of each brain region were calculated to explore the key brain regions influencing the brain age prediction results. Evaluation was carried out on two public datasets， E-NKI and Cam-CAN. It can be seen that the memory required for LMCN parameters is 2.30 MB， which is 60.3% and 52.0% less than those of MobileNetV3 and ShuffleNetV2 respectively. In terms of prediction results， on E-NKI dataset， LMCN has the Mean Absolute Error （MAE） of 5.16， the R of 0.947， and the Root Mean Square Error （RMSE） of 6.40. Compared to the model that combines network-based feature selection with the least angle regression， LMCN has the MAE decreased by 1.34 and the R increased by 0.037； on Cam-CAN dataset， LMCN has the MAE of 5.97， the R of 0.904， and the RMSE of 7.93. Compared to the connectome-based machine learning model， LMCN has the R increased by 0.019， and the RMSE decreased by 0.64. The results show that while LMCN has small number of parameters and is easy to deploy， it can improve the accuracy of fMRI brain age prediction effectively and provide clues for assessing the brain status of healthy adults.

Key words: functional Magnetic Resonance Imaging (fMRI), Functional Connectivity (FC), lightweight neural network, brain age prediction, deep learning

摘要：

针对功能磁共振成像（fMRI）脑龄预测精度较低以及该问题与深度学习结合研究较少的现状，提出一种轻量级多尺度卷积网络的fMRI脑龄预测模型（LMCN）。首先，通过计算fMRI中感兴趣区（ROI）的皮尔逊相关系数（R）得到ROI的功能连接（FC）矩阵，并将该矩阵作为输入；其次，在提升FC通道数保证特征数的同时缩小特征图尺寸，并采用具有人类视觉注意力特点的多尺度空洞卷积模块RFB（Receptive Field Block）提取年龄特征；最后，通过全连接层输出预测脑龄，并计算各脑区的消融预测结果，从而探索对脑龄预测结果产生影响的关键脑区。在E-NKI和Cam-CAN这2个公开数据集上评估，可知LMCN参数所需内存为2.30 MB，比MobileNetV3、ShuffleNetV2分别减少了60.3%、52.0%。预测结果方面，在E-NKI数据集上，LMCN的平均绝对误差（MAE）为5.16，R为0.947，均方根误差（RMSE）为6.40，与基于网络的特征选择结合最小角回归模型相比，MAE减小了1.34，R增加了0.037；在Cam-CAN数据集上，LMCN的MAE为5.97，R为0.904，RMSE为7.93，与基于连接组的机器学习模型相比，R提升了0.019，RMSE减小了0.64。结果表明，LMCN在参数量较小易于部署的同时，能够有效提高fMRI脑龄预测的精度，并为评估健康成人的脑部状态提供线索。

关键词: 功能磁共振成像, 功能连接, 轻量级神经网络, 脑龄预测, 深度学习