基于多模信息的帕金森症的诊断和预测方法研究

发布时间：2018-03-12 07:36

本文选题：帕金森症　切入点：特征选择　出处：《深圳大学》2017年硕士论文　论文类型：学位论文

【摘要】：帕金森是世界上第二大的神经衰退性疾病。该疾病往往伴有多种非运动型症状(例如:抑郁、嗅觉障碍、睡眠紊乱等),运动症状包括迟缓、颤抖、僵硬等,此外,尽管这些症状是由多巴胺神经元的减少造成,但当今仍未发现这些多巴胺神经元减少的具体原因,因此诊断和治疗通常是复杂的。当前已有一些机器学习的方法运用在帕金森诊断和临床得分预测上,结果表明能有效的预测不同阶段帕金森症,即帕金森进程跟踪。这些方法大都关注于使用单模态的数据完成单功能模式,即在同一个方法模型中,用某一种单一模态数据来完成分类功能或者仅完成预测功能。一般来说,分类和回归模型有3个关键步骤:特征提取、特征选择、分类或回归。在医学图像研究中,从样本数量少而信息维度高的特征中选择出相关性最大的特征一直是个棘手的问题。因此,我们提出了一种基于改进损失函数的统一的帕金森症特征选择模型,该模型通过使用多模态帕金森影像学数据,同时完成了帕金森症诊断和预测功能。该模型用一种改进的损失函数来抓取那些对自身模型有效的特征,然后在这些特征上运用支持向量回归模型和支持向量分类模型来分别达到得分预测和症状检测的目的。我们用了10次交叉验证的方法对208个被试进行实验,208个被试中包括56个正常组、123帕金森组以及29特别组(scans without evidence of dopamine deficit,SWEDD)。实验结果表明,与只使用单个模态的数据相比,多模态数据能够有效的提高帕金森检测和得分预测性能。实验对比分析得到,我们的方法要比当前主流的方法更有效。
[Abstract]:Parkinson's is the second largest neurodegenerative disease in the world. The disease is often accompanied by a variety of non-motor symptoms (e.g. depression, olfactory disorders, sleep disorders, etc.). Motor symptoms include retardation, tremors, stiffness, etc. Although these symptoms are caused by a decrease in dopamine neurons, no specific cause of these decreases has yet been found. Therefore, diagnosis and treatment are usually complicated. Currently, some machine learning methods have been used in Parkinson's diagnosis and clinical score prediction, and the results show that they can effectively predict different stages of Parkinson's disease. These methods mostly focus on using single mode data to complete single function pattern, that is, in the same method model, one single mode data is used to complete classification function or only to complete prediction function. Classification and regression models have three key steps: feature extraction, feature selection, classification or regression. It has always been a difficult problem to select the most correlated feature from the features with small sample size and high information dimension. Therefore, we propose a unified feature selection model for Parkinson's disease based on improved loss function. The model uses multimodal Parkinson imaging data to perform both diagnostic and predictive functions for Parkinson's disease. It uses an improved loss function to capture features that are valid for its own model. Then support vector regression model and support vector classification model are used to predict the score and detect symptoms on these features. We used 10 cross-validation methods to carry out experiments on 208 subjects. The subjects included 56 normal subjects, including Parkinsonian and 29 special groups. The results showed that, Compared with single modal data, multimodal data can effectively improve the performance of Parkinson's detection and score prediction. The experimental results show that our method is more effective than the current mainstream methods.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R742.5;TP391.41

【相似文献】