Research On Early Diagnosis Of Alzheimer’s Disease Based On Deep Learning

Alzheimer’s disease(AD)is an irreversible neurodegenerative disease that seriously endangers the physical and mental health of the elderly.With the aging of the population,the incidence of AD in our country is still increasing.Due to the slow onset of AD,the patients have reached the late stage as soon as the symptom is discovered,and there is no effective curative treatment.However,if AD can be detected as early as possible,the progress of AD can be effectively slowed down or even inhibited through implementing treatment and intervention.Therefore,realizing the early diagnosis of AD has become an important research topic in the field of computer-aided diagnosis.Mild cognitive impairment(MCI)is a prodromal stage of AD.Accurate identification of MCI is of great significance for the early diagnosis of AD.As a pathological biomarker,positron emission tomography(PET)has been widely used in the diagnosis of AD and MCI,and has significant advantages in the early diagnosis of AD.Therefore,PET is used to achieve AD diagnosis in our work.At present,most related researches are dedicated to the study of the binary classification of AD and normal controls(NC)or MCI and NC.However,from the perspective of practical application,an Alzheimer’s disease system needs to have the ability to accurately identify samples of AD,MCI and NC categories,so it is very necessary to build a robust three-category model.Due to the difficulty of classification between MCI and AD,NC,it is often not able to achieve an optimal performance by directly designing a 3-category classifier.In order to solve this problem,the three-category classification problem is divided into three binary classifications.The outputs of the three binary classification networks are fused to achieve the final classification of AD,MCI and NC.For each binary classification problem,this paper proposes a joint model based on CNN and SVM.In view of the powerful feature extraction ability of CNN,a robust 3DCNN model is built to extract image features.Since SVM can obtain the optimal solution in binary problems and has advantages in small sample problems,SVM is used to classify the features extracted by 3DCNN.Due to the different optimization methods of multi-layer perceptron and SVM,end-to-end training cannot be realized between 3DCNN and SVM.In order to solve this problem,this paper proposes an end-to-end training strategy based on 3DCNN+SVM,which realizes the overall optimization of 3DCNN+SVM and improves the prediction performance of the model.Python language and Pytorch framework are used to implement the entire model.The data used in our work comes from the ADNI database,which contains 2706 PET images from267 AD subjects,340 MCI subjects and 352 NC subjects.A large number of experiments based our algorithms show that,compared with other papers,our algorithms can achieve the best performance.At the same time,a prototype application system based on all the algorithms designed in this article is developed for AD diagnosis,which lays foundation for further in-depth research and application.