Research On Parkinson's Disease Classification Algorithm Based On Feature Fusion And Convolutional Neural Networ

Parkinson’s disease(PD)ranks as the second most prevalent neurodegenerative disorder worldwide,primarily caused by the loss of nigrostriatal cells in the brain.The disease manifests itself in symptoms such as movement disorders and tremors,which can significantly impact patients.However,early-stage Parkinson’s patients often do not present clinical symptoms,making it challenging to diagnose and resulting in missed treatment opportunities.Therefore,an early diagnosis of PD is crucial.Magnetic resonance imaging(MRI)is a medical imaging technique that reveals structural changes in the brains of people with Parkinson’s disease.In the context of medical image classification,feature fusion can extract richer and more comprehensive image features,leading to improved classification performance.Based on MRI images,this thesis studies the Parkinson’s disease classification algorithm based on feature fusion and convolutional neural network:(1)A classification algorithm for Parkinson’s disease based on the fusion of texture features and improved ResNet18 features is proposed.First,the 2D MRI image dataset required in this thesis is constructed for the current PD classification problem.Three slices around the center of each subject volume were picked by slicing the 3D MRI images of the subjects.Secondly,through the traditional heuristic methods of LBP and GLCM,the texture features of the image after data enhancement are extracted.Improve the original ResNet18 network and add a feature compression layer to extract the deep features of the image.Finally,after fusing the texture features and deep features,the support vector machine was used for five-fold cross-validation,and a classification accuracy of 94.26% was obtained.(2)A Parkinson’s disease classification algorithm based on automatic traditional feature extraction and gray wolf optimized ResNet34 feature fusion is proposed.First of all,the automatic traditional feature extraction algorithm(Automatic Traditional Feature Extraction,ATFE)is proposed to use five different feature extraction methods to automatically extract the edge,texture and other features of the image.By splicing the extracted features,each image obtains 208-dimensional features.vector.Secondly,the hyperparameters of SGD during the ResNet34 training process are optimized through the Gray Wolf Optimizer(GWO),which improves the efficiency and accuracy of parameter tuning and enhances the feature extraction capability of the network.Finally,the features extracted by ATFE were fused with the features extracted by gray wolf optimized ResNet34,and five-fold cross-validation was performed through random forest.The classification accuracy was 95.93%,and the sensitivity and specificity were 95.63% and 97.46%,respectively.(3)A Parkinson’s disease classification algorithm based on feature fusion of shallow convolutional neural network and deformable ResNet50 is proposed.First,a shallow convolutional neural network model,called PD-CNN,is constructed to automatically extract shallow features of images.Secondly,improve the ResNet50 network,introduce the DCNv2 deformable convolution module,and construct the DCNv2-ResNet50 network.The network can extract deep feature representations based on lesion shape.Finally,the shallow features are fused with the deep features,and the support vector machine,K-nearest neighbor algorithm,random forest,decision tree and naive Bayesian are used for classification.The support vector machine achieves the best classification accuracy rate of 98.64%.In addition,the features extracted by PD-CNN and DCNv2-ResNet50 are respectively subjected to five-fold cross-validation using SVM,and the results demonstrate the effectiveness of the proposed network.This thesis proposes the feature fusion PD classification algorithm,which solves the problem of low classification accuracy of Parkinson’s MRI images using a single feature,and is of great significance for the diagnosis of Parkinson’s disease.