Research On Small Sample Learning Method For Auxiliary Diagnosis Of Neurodegenerative Diseases

Neurodegenerative diseases(NDDs)threaten the health of the elderly at all times,and even endanger the lives of patients in severe cases.In our country,due to the low doctor-patient ratio,the poor consciousness of the target population for prophylaxis and treatment,and the good stealthiness of the disease itself,NDDs have a low diagnosis rate,high misdiagnosis rate and low treatment rate.Machine learning methods can improve the efficiency and accuracy of NDDs diagnosis by intelligently analyzing images,speech and other data related to NDDs.However,because NDDs have the characteristics of strong stealthiness and large differences between different people,the data has the characteristics of high aliasing(large intra-class divergence and small interclass distance)and small samples,which brings many outstanding issues to the intelligent diagnosis of NDDs by machine learning methods.These issues include: 1)High aliasing and small sample characteristics of NDDs data easily lead to low recognition accuracy of existing machine learning methods;2)Due to the high aliasing characteristics of NDDs data,the distribution matching between data does not necessarily lead to high classification accuracy.However,the existing homogeneous domain adaptive methods are oriented by distribution matching,which hinders the performance improvement of the methods;3)Disease data in different periods have feature spatial heterogeneity,which is mainly manifested in the source domain features are less or far less than the target domain features,making the domain adaptive method not directly applicable.These issues have brought great challenges to the auxiliary diagnosis of NDDs.Aming at the above three scientific issues,with the support of the National Natural Science Foundation of China,the Central University Fund and the Graduate Research and Innovation Fund of Chongqing,there are two most common NDDs-Alzheimer’s disease(AD)and Parkinson’s disease(PD)has been researched in this thesis.Starting with dimensionality reduction and domain adaptation,research work is carried out in three aspects: local intra-class decision-preserving projection,supervised cross-domain adaptation and heterogeneous domain adaptation.The main research work and innovations of this thesis are as follows:(1)Noise-added integrated local intra-class decision-preserving projection method for diagnosis of NDDsDue to the high aliasing and small sample characteristics of NDDs data,there is insufficient inter class discrimination of disease data and "over fitting" phenomenon of the model,which makes the low accuracy of disease diagnosis.To solve this issue,a noisy-added integrated local intra-class decision-preserving projection method is proposed in this thesis.Firstly,aiming at the small sample characteristics of disease data,the proposed method achieves the effect of regularization through micro-noise embedding,which effectively alleviates the "overfitting" of the model;Then,according to the high-aliasing characteristics of the data,a decision preserving projection method is designed.By using the label information of data,the priority is given to minimizing the within class divergence of data,increasing the inter class distance of data and maintaining the local characteristics of data decisively,so that the discrimination between data categories can be improved;Finally,the diagnostic performance of the model is further improved by introducing the Bayesian fusion strategy to construct the mapping matrix.Experimental results on multiple AD magnetic resonance images(MRI)and Parkinson’s speech datasets shown that the proposed algorithm can achieve higher diagnostic accuracy compared with existing dimensionality reduction methods used in NDDs.(2)Cross-domain adaptive method of Fisher’s discriminant criterion for diagnosis of NDDsThe high-aliasing characteristics of NDDs restrict the improvement of the disease diagnosis performance of the distribution matching-oriented domain adaptive method.To address this issue,a domain adaptive method based on Fisher’s discriminant criterion,which is oriented by the degree of discrimination between classes,is proposed in this thesis.This method increases the distance between samples of different categories in the target domain and at the same time uses the labeled samples in the target domain as a guide to achieve the clustering of the source domain samples to their corresponding target domain class centers,which makes the source domain and target domain data after mapping have a higher degree of discrimination between classes.In addition,since the proposed method has a closed-form solution,it has a lower time cost than most domain adaptive methods based on distribution matching.Experimental results on multiple PD speech datasets shown that the proposed method has higher accuracy and lower time consumption than the existing domain adaptive methods used in PD speech diagnosis.(3)Source-domain few-feature cross-domain cross-manifold metric alignment method for diagnosis of NDDsThe gradual nature of cognition and other factors make the data of NDDs in different periods have characteristic spatial heterogeneity,which makes the domain adaptive methods not directly applicable.Aiming at its specific manifestations-the source domain features are less or far less than the target domain features,a heterogeneous domain adaptive method-cross-domain cross-manifold metric alignment method based on few features of the source domain is proposed in this thesis.Firstly,aiming at the few feature characteristics of source disease data,the method realizes the consistency of the source domain and target domain feature spaces through the proposed nearest neighbor interactive feature growth algorithm;Then,according to the category,the source domain and target domain data are geometrically aligned through the proposed cross-domain cross-manifold metric alignment method.While realizing the migration of the data from the source domain to the target domain,it reduces the degree of aliasing of disease data through the way of manifold metric alignment;Finally,the experimental results on different periods of AD MRI data shown that the feature growth algorithm based on nearest neighbor interaction can effectively achieve the consistency of the feature space of the disease data in the source domain and the target domain,and the proposed cross manifold metric alignment method has higher diagnostic accuracy than the existing domain adaptive methods.