毕业生知识库

精神分裂症是一种常见、高致残的重性精神类疾病，精神分裂症的发病机制
不明，现有的精神分裂症诊断主要根据主观的现象观察和家人叙述，没有一个客
观的事实依据，给精神分裂症的诊断和治疗带来了极大的不便。为了给临床提供
客观的依据和高效的辅助诊断工具，减轻医生的诊断难度，给医生提供客观的角
度。本文基于机器学习算法，首次在多中心、大样本的精神分裂症以及健康对照
组的结构核磁共振数据(sMRI)上进行基于体素和感兴趣区域的模型构建和结果分
析。此外，本文还重点研究了模型的鲁棒性并且发现了精神分裂症相关的潜在生
物学标记以及对临床数据做了相关性分析。

首先基于采集的八个独立站点的结构核磁共振数据集，总共1275 个被试数
据，包含613 个健康对照组，以及662 个精神分裂症患者。对数据提取结构相关
的灰质体积、白质体积、脑脊液体积作为训练数据的特征。为了降低模型的复杂
度，首先采用皮尔逊相关系数在训练数据集合上进行维度约减后，建立深度神经
网络模型，采用留一站点法进行结果验证，分类结果在独立站点上可以重复验证，

最终在独立站点测试集上可以达到80.20%的平均准确率，达到目前最好的性能。此外，对比于传统的机器学习方法，如支持向量机、随机森林，分类准确率均有很明显的提升。

其次，对于临床转化来说，模型的决策过程对临床医生来说尤其重要。本文
结合逐层相关度传播算法，分析个体被试数据的特征在模型决策过程中的作用。
为了验证深度学习模型得到的精神分裂症相关特征的可靠性，本文分别从荟萃分
析结果和线性模型两个角度进行验证。多次比较发现，和精神分裂症相关的结构
异常脑区主要集中在颞上回、岛叶皮层和颞叶等区域。此外，我们分析了不同种
类的精神分裂症（首发病人和复发病人）在模型分类上的差异以及特征上的差异
性，其中发现复发病人在分类精度上稍高于首发病人。此外，我们分析了特征数
据和用药信息、一般功能评分以及PANSS评分之间的相关性分析。
然后，基于体素水平的高维特征会给临床转化的可解释性和使用带来困难，
因此本文进一步基于精细的脑网络组图谱提取感兴趣区域（Region of Interest, ROI）的210个脑区的皮层厚度、表面积、体积和61个脑区的子结构体积。相较于原始的体素水平的特征，感兴趣区域的数据特征十分简单、易于解释。本文提出多站点深度学习模型，充分考虑不同站点在特征学习过程中的差异性，我们的多站点深度学习模型比传统的深度学习模型（即多层感知器）在分类准确率上提高了2.2%。同时，从感兴趣区域的特征角度验证深度学习模型的可解释性，发现重要的特征主要有颞上回、岛叶和扣带等区域，和体素水平的结论基本一致。

总之，本文对精神分裂症和正常人进行了全面和多角度的分析，分别从感兴趣区域和体素水平两个特征维度出发，使用机器学习模型在多中心大样本数据集
上对精神分裂症和正常人进行模式分类。并验证了模型结果的可解释性以及对临
床数据进行相关性分析。

Schizophrenia is a common, highly disabling and severe mental illness. The pathogenesis of schizophrenia is unknown. The existing diagnosis of schizophrenia is mainly based on subjective observations and family narratives. There is no objective basis, it is difficult to diagnose and treat schizophrenia, in order to provide objective basis and effective diagnostic tools to ease the diagnosis of doctors, Based on machine learning algorithms, this paper focus on in a multi-center, large sample of schizophrenia patients and health controls with structural nuclear magnetic resonance data (sMRI) on voxel level and
Region-of-Interest(ROI) level for the first time, as well as the robustness of the proposed model and the discovery of schizophrenia relevant potential biomarkers and correlations with clinical data.

Firstly, this paper collected structural MRI datasets of eight independent sites, a total of 1275 test data, which contains 613 healthy controls, as well as 662 schizophrenia patients, gray matter associated with data extraction structures volume, white matter volume, cerebrospinal fluid volume as a feature of training data, in order to reduce the complexity of the training data, after using the Pearson correlation coefficient to reduce the dimension on the training data set, then the deep neural network model is established, and the result is verified by the leave-one-site-out method. The classification result can be verified repeatedly on the independent site. Finally, the average value of the independent site test set is 80.20%. which can achive state-of-the-art perfromance. Besides. Compared to traditional machine learning methods, Such as SupportVector Machine(SVM), Random Forest(RF), the classification accuracy rate has been significantly improved.

Secondly, for clinical transformation, the decision-making process of the model is especially important for clinicians. This paper combines the layer-wise relevance propagation algorithm. Analyze the role of individual subject data in the model decision process. To verify the reliability of the schizophrenia-related features obtained by the deep learning model, This paper verifies the results from the meta-analysis and the linear model. It has been found that the structural abnormalities associated with schizophrenia are mainly concentrated in the superior temporal gyrus, insular cortex and temporal lobe. In addition, different types of schizophrenia (first episode and relapsed patients) were analyzed.Differences in model classification and differences in features. Among them, it was found that patients with only relapse patients had a slightly higher classification accuracy than those with only the first episode patients.

Thridly, this paper combines the fine-grained Brainnetome Atlas to extract featuresAnd This paper propose the generalized feature-invariant deep neural network framework to ensure the model generalization in automatic diagnosis schizophrenia. Our model evaluate with 10-fold cross-validation method and leave-one-site validation method, the average accuracy brings 2.2% gain in classification accuracy against standard leave-one-site-out validation. Besides, our model gives the promising result on all sites classification and great potentials for computer-aided diagnosis of psychiatric disorders with simple and meaningful biomarkers.

All in all, this paper analyze schizophrenia patients and normal controls from different perspective, respectively from ROI-level and voxle-level two characteristics using machine learning model in multisite large sample data sets for schizophrenia and normal pattern classification. Besides, this paper analyze and explain the results of the model, the analyze
the correlation between feature and clinical data.