【作者】 王涛；

【导师】 黄美燕；

【作者基本信息】 南方医科大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 神经退行性疾病(Neurodegenerative Diseases,NDs),如阿尔兹海默病(Alzheimer’s Disease,AD)和帕金森病(Parkinson’s Disease,PD),是一类不可逆的进行性疾病,及早预测NDs的发生可提高诊疗效率,具有重大临床意义。随着成像和基因技术的发展,产生的多模态影像数据和遗传数据可为NDs预测提供互补信息。NDs作为进行性疾病,随访获得的纵向数据中包含的疾病进展信息可为预测提供依据。通过深度学习模型利用多模态和纵向数据实现疾病预测可为临床诊疗提供辅助。但是,多模态数据的模态缺失和纵向数据的随访缺失问题阻碍着这些数据的有效应用。在保留包含缺失信息的数据的前提下,如何保证模型合理利用所有数据是有效提升预测性能的核心问题。本文首先探究影像数据的缺失数据处理和疾病预测方法,然后进一步结合遗传数据展开分析与探索。神经影像数据可反映大脑结构和功能信息,也是采用深度学习方法构建NDs预测模型中最常用的数据类型,我们首先聚焦于基于神经影像数据的NDs预测模型。在临床实践中,NDs预测通常使用单时间点数据,尤其是基线访问(Baseline visit,BL)数据。因此,模型在实际使用时应在没有纵向数据的介入下完成预测。本文提出了多角度插补和交叉注意力网络将缺失数据插补和疾病预测整合在一个框架内,并在BL处完成AD的准确预测:首先,利用一种结合对抗学习的多角度插补方法以较小的误差处理各类缺失数据;其次,引入交叉注意力策略挖掘数据中不同角度的潜在关联以辅助预测;通过适当的训练,网络可利用纵向数据中的疾病进展信息和多模态数据中的互补信息提高在BL处的预测能力。遗传因素与NDs的发生和进展也息息相关,后续考虑将神经影像数据和遗传数据结合进行NDs预测,以影像遗传学手段建立深度学习模型。以往的方法多采用先融合多模态影像数据再与遗传数据进行关联的策略,导致难以有效挖掘多模态影像数据中的共有和互补信息,且未能准确表达影像与遗传数据之间的复杂关系。本文提出了深度多模态解耦关联分析网络来解决数据缺失及上述问题:首先,引入多模态解耦模块解耦经过非线性投影的影像表征,将不同模态影像表征分为模态共有和特有部分;其次,遗传数据也以非线性的方式分别映射至解耦后的影像表征上,建立影像和遗传数据间的联系,并同步生成掩码向量以辅助后续AD和PD预测;当影像数据缺失时,映射的遗传表征可替代缺失模态的影像表征,实现对包含缺失信息数据的有效利用。所提出的方法均在独立测试集上进行了测试,其性能优于现有先进方法,验证了提出方法的有效性。此外,通过模型对NDs生物标志物进行了检测,保证了模型的可解释性,并为进一步探究NDs的病理机制提供帮助。更多还原

【Abstract】 Neurodegenerative diseases(NDs),such as Alzheimer’s Disease(AD)and Parkinson’s Disease(PD),are a kind of chronic progressive diseases.Moreover,early prediction of NDs can be used to improve the diagnostic and therapeutic efficiency,which is of great clinical significance.With the development of imaging and genetic technologies,increasing multimodal imaging and genetic data can provide complementary information for NDs prediction from multiple perspectives.Since NDs are progressive diseases,disease progression information inherent in the longitudinal data from follow-ups can also provide a basis for NDs prediction.Besides,deep learning methods for disease prediction based on longitudinal and multimodal data can provide assistance for clinical diagnosis and therapy.However,missing modality issues in multimodal data and missing follow-up issues in longitudinal data hinder the effective application of multimodal and longitudinal data.On the premise of retaining data with missing information,it is critical to ensure that all available data can be reasonably applied for improvement of prediction performance.In this thesis,missing data problem exist in imaging data are first explored to achieve accurate disease prediction accuracy,and then,genetic data are included to further improve performance of disease prediction.Neuroimaging data can reflect varied information of brain structure and function,which are the most commonly used data type in deep learning methods for NDs prediction.Thus,we firstly explore methods of NDs prediction based on neuroimaging data.In clinical practice,previous studies focused on using single time point data for NDs prediction,especially data at the baseline visit(BL).Therefore,NDs prediction should be achieved without the intervention of longitudinal data in practical use.In this thesis,a multi-view imputation and cross-attention network is proposed to integrate missing data imputation and disease prediction in a unified framework,and accurate prediction of AD can be achieved by using BL data.First,a multi-view imputation method combined with adversarial learning is utilized to handle various types of missing data with small errors.Second,two cross-attention blocks are introduced to explore potential associations in longitudinal and multimodal data from different views to assist in prediction.With proper training,the disease progression information in longitudinal data and the complementary information in multimodal data can be exploited to improve the prediction performance at BL.Genetic data are closely related to the occurrence and progression of NDs.Therefore,neuroimaging and genetic data can be combined,generally denoted as imaging genetics,by using a deep learning method for NDs prediction.However,previous methods firstly fused multimodal data and then correlated them with genetic data,which failed to explore common and complementary information among multimodal imaging data and construct complex relationships between imaging and genetic data.In this thesis,a deep multimodality-disentangled association analysis network is proposed to solve missing data issue and aforementioned problems.First,a multimodality-disentangled module is introduced to disentangle the imaging representations with nonlinear projection.Moreover,the imaging representations of different modalities are divided into modality-common and modality-specific parts.Second,the genetic data are mapped to the disentangled imaging representations in a non-linear manner.Meanwhile,the associations between the imaging and genetic data are built,and mask vectors are synchronously generated to assist subsequent prediction of AD and PD.Moreover,the mapped genetic representations can be used to replace the imaging representations of the missing modality to handle missing data problem.The performance of the proposed methods are evaluated on independent testing sets,and the experimental results show that the proposed methods outperform several competitive methods,which verifies the effectiveness of the proposed methods.Moreover,NDs related biomarkers are detected via the proposed methods to ensure the interpretability of the proposed methods.Therefore,the proposed methods can provide potential tools to give an insight for the pathological mechanisms of NDs.更多还原