基于深度主成分相关自编码器的多模态影像遗传数据研究

脑成像表型和基因变异已成为影响精神分裂症等复杂疾病的重要因素。研究人员根据以往在致病机理方面的深入研究,已经提出了很多基于深度神经网络或正则化的模型,这些模型通常包含某种形式的惩罚项或具有重建目标的自编码器结构,但其所使用的多模态数据的特征维数往往大于样本个数。为了应对高维数据分析的困难并突破深度典型关联分析的局限性,文中提出了一种由多模态线性特征学习的主成分分析和基于限制玻尔兹曼机的多模态非线性特征学习的多层信念网络组成的有效模型。该模型和先前的先进模型一起被应用在实际的多模态数据集上进行测试和分析。实验发现,与已有模型相比,深度主成分相关自编码器模型学习的特征具有更高的分类性能和更强的关联性。在分类精度方面,两类模态数据的分类精度均超过了90%,相比平均精度在65%左右的基于CCA的模型和平均精度在80%左右的基于DNN的模型,该模型的分类效果有了显著提高。在聚类性能评估的实验中,该模型以93.75%的平均归一化互信息指标和3.8%的平均分类错误率指标进一步验证了其优越的分类性能。在最大关联性分析方面,当顶层节点输出维度一致时,该模型以0.926的最大关联性胜于其他先进模型,在高维数据分析方面表现出了优异的性能。

Study on Multimodal Image Genetic Data Based on Deep Principal Correlated Auto-encoders

Brain imaging phenotype and genetic mutation has become the important factors that affect complex diseases such as schizophrenia,researchers based on previous work in the pathogenesis of in-depth research have proposed many models based on deep neural network or regularization,typically involving either some form of norm or auto-encoders with a reconstruction objective,but the multi-modal data of those models tend to have the number of feature dimensions which more than that of samples.In order to solve the difficulties of high-dimensional data analysis and overcome the limitations of deep canonical correlation analysis,a competent optimization algorithm is exploited to solve deep canonical correlation analysis(DCCA)with principal component analysis(PCA)on the multi-modal linear features learning and multi-layer belief network based on restricted Boltzmann machine(RBM)on multi-modal nonlinear features learning.The model,together with previous advanced model,has been applied to test and analyze the actual multi-modal data.Experiments show that the deep principal component correlation auto-encoders model has higher correlation and better classification performance than those previous model.In terms of classification accuracy,the classification accuracy of the two types of modal data is more than 90%.Compared with the CCA-based model with an average accuracy of about 65%and the DNN-based model with an average accuracy of about 80%,the classification effect of this model is significantly improved.In the experiment of clustering performance evaluation,the model further verified the significant classification effect of the model with average normalized mutual information of 93.75%and average classification error rate of 3.8%.In terms of maximum correlation analysis,on the premise that the output dimensions of top-level nodes are consistent,this model outperforms other advanced models with the maximum correlation of 0.926,showing excellent performance in high-dimensional data analysis.