结构磁共振成像（structural magnetic resonance imaging, sMRI）可以测量脑组织成分的变化信息，是一种常用的神经影像技术。基于sMRI的机器学习方法被广泛应用于神经系统相关疾病的分类预测研究中，为疾病的早期诊断及预防提供了重要依据。由于深度学习的飞速发展，神经网络也越来越受到研究者的重视，其中，Siamese网络由于网络分支共享权值的特性，在纵向数据建模中有独特的优势，由于纵向数据能够将神经影像在时间维度上的变化信息融入模型，从而有助于获取疾病发展进程中更全面的信息。轻度认知障碍（Mild Cognitive Impairment, MCI）是阿尔茨海默氏病（Alzheimer’s Disease, AD）发展的前驱阶段，每年约有15％的MCI患者转化为AD，因此对MCI的转化进行有效预测，对AD的早期诊断及临床干预具有重要意义。

本研究针对Alzheimer’s Disease Neuroimaging Initiative（ADNI）数据集的281名MCI被试的纵向sMRI数据与临床特征，建立了基于Siamese网络的预测模型，实现了对MCI转化以及转化时间的预测。首先，以两个三维卷积神经网络（Convolutional Neural Network, CNN）作为Siamese网络分支，提取纵向sMRI特征。通过纵向差异编码（Longitudinal Difference Coding, LDC）整合两分支特征，在其后接全连接层，并加入临床特征，预测MCI转化，通过五折交叉验证评估模型性能。其次，为评估纵向数据的预测性能，本文建立了单分支Siamese网络预测模型，即基于横向数据的CNN模型；此外，为评估Siamese网络的预测性能，本文建立了基于纵向数据的支持向量机（Support Vector Machine, SVM）和随机森林（Random Forest, RF）模型，进行对比分析。最后，将预测MCI转化的最优Siamese网络模型作为预训练模型，在其上进行微调，建立了预测MCI转化者（MCI converter, MCI-c）转化时间的模型，并与无预训练的Siamese网络进行对比，通过五折交叉验证评估模型性能。

基于纵向数据的Siamese网络预测MCI转化的结果显示，使用sMRI数据结合临床特征的预测效果最好，准确率达到81.25%，而基于横向数据的CNN的预测准确率为79.53%，低于纵向Siamese模型。基于纵向数据的SVM和RF的预测准确率分别为76.87%和76.17%。上述结果表明，基于纵向数据的Siamese网络在预测MCI转化中有一定的优势。基于Siamese网络预测MCI-c转化时间的结果显示，基于预训练的Siamese网络的预测准确率为70.58%，无预训练的Siamese网络的预测准确率为58.74%。表明通过预训练，能够更好地提取出与MCI转化相关的sMRI特征，从而得到更好的预测结果。基于预训练的Siamese网络在独立测试集上预测准确率为68.09%，与验证集上表现接近，证明了模型的有效性。

本研究证明了CNN具有强大的特征提取能力，同时，Siamese网络能够很好地结合纵向数据。以三维CNN作为Siamese网络分支提取sMRI特征，同时结合临床指标，能够较准确地预测MCI的转化并在此基础上预测MCI-c的转化时间，对于AD的早期诊断具有一定的临床价值。

Structural magnetic resonance imaging (sMRI) can measure changes in brain tissue composition and is a commonly used neuroimaging technique. Machine learning methods based on sMRI are widely used in the classification and prediction of neurological diseases, providing an important basis for the early diagnosis and prevention of diseases. Due to the rapid development of deep learning, neural networks have attracted more and more attention from researchers. Among them, Siamese Network (SN) has a unique advantage in longitudinal data modeling due to the characteristics of sharing weights of network branches, which allows to integrate the time-related change information into the model. As a result, it helps to obtain more comprehensive information in the course of disease development. Mild Cognitive Impairment (MCI) is a prodromal stage of Alzheimer's Disease (AD). The conversion rate of MCI to AD is about 15% per year. Efficient prediction of MCI conversion is of great significance for the early diagnosis and clinical intervention of AD.

In this study, we used longitudinal sMRI and clinical attributes of 281 MCI subjects from Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset to establish a prediction model based on SN to predict the conversion of MCI and the conversion time of MCI-c as well. Firstly, we took two 3D convolutional neural networks (CNN) as branches of SN to extract longitudinal sMRI features. Longitudinal Difference Coding (LDC) was used to integrate the two branch features, followed by fully connected layers, in which the clinical features were added to predict MCI conversion. The model performance was evaluated by 5-fold cross-validation. Secondly, in order to evaluate the prediction performance of longitudinal data, this paper established a single branch SN prediction model, that is, a CNN model based on cross-sectional data. In addition, in order to evaluate the prediction performance of SN, this paper established a support vector machine (SVM) model and a Random Forest (RF) model based on longitudinal data for comparative analysis. Finally, The optimal MCI conversion predicting model was used as a pretraining model, on which fine-tuning was performed to establish a model for predicting MCI-c conversion time. This model was called SN based on pre-training and was compared with SN without pre-training. The model performance was evaluated by 5-fold cross-validation.

The prediction results of SN based on longitudinal data shows that the use of sMRI data combined with clinical attributes has the best prediction performance, with an accuracy rate of 81.25%. The accuracy of CNN based on cross-sectional data is 79.53%, which is lower than that of SN based on longitudinal data. The prediction accuracy of SVM and RF based on longitudinal data is 76.87% and 76.17%, respectively. The above results show that SN based on longitudinal data has certain advantages in predicting MCI conversion. The prediction results of the MCI-c conversion time based on SN show that the prediction accuracy of the pre-trained SN is 70.58%, and the prediction accuracy of SN without pre-training is 58.74%. It shows that through pre-training, the sMRI features related to MCI conversion can be better extracted, thereby obtaining better prediction results. The prediction accuracy of SN based on pre-training on the independent test set is 68.09%, which is close to the performance on the validation set, proving the effectiveness of the model.

This study shows that CNN has a powerful feature extraction capability and SN can effectively combine longitudinal data information. Using 3D CNN as a branch of SN to extract sMRI features, combined with clinical attributes, can accurately predict the conversion of MCI and predict the conversion time of MCI-c on this basis, which has clinical value for the early diagnosis of AD to some extent.