Model proposal of demxnet for the diagnosis of Alzheimer's disease and comparison with deep transfer learning methods

Dementia is a syndrome in which cognitive functions are impaired beyond the usual consequences of biological aging. The most common type of dementia is Alzheimer's disease. It exacerbates before visible symptoms occur and cause an incurable brain condition. Therefore, the sooner the diagnosis is made, the more preventive activities and treatment can be started for the patient. Diagnosis is based on magnetic resonance imaging (MRI) brain analysis, although tests such as the Mini-Mental Status Exam are often used for early diagnosis. In this thesis, the performances of deep learning models for the diagnosis and classification of Alzheimer's dementia disease are compared, and a new deep learning-based model called DEMXNET is proposed. The dataset of MRI images containing four classes of Alzheimer's disease was obtained from the Kaggle platform. Due to the uneven distribution of the number of images in the classes in the data set, Synthetic Minority Oversampling Technique (SMOTE) was used. All models were trained and tested in two scenarios, with and without SMOTE. It was observed that the accuracy rates of the SMOTE technique increased significantly in all models. Accuracy rate performance results of the models with the SMOTE technique used in the classification of Alzheimer's type dementia disease, VGG16 (92.89%), VGG19 (95.94%), InceptionV3 (89.65%), Xception (91.05%), DenseNet201 (95.47%) and our proposed model DEMXNET for (96.48%). The proposed DEMXNET model was more successful than the other models compared. The proposed DEMXNET model can be used efficiently, especially for unbalanced datasets, and can be adapted to different datasets.