Classification and rating of Alzheimer's disease by using brain images and clinical features with hybrid deep learning methods

Early diagnosis and identification of the stages of Alzheimer's disease is of great importance to slow down the progression of the disease and improve the quality of life of patients. This thesis aims to improve Alzheimer's diagnosis and classification accuracy using machine learning and deep learning methods. The study utilizes MRI images and biomarker data from OASIS and ADNI datasets; the sample consists of Alzheimer's, MCI and healthy individuals. First, a classification system was developed with tree-based algorithms (Decision Tree, Random Forest, Gradient Boosting, etc.) and the Gradient Boosting algorithm gave the most successful results with an accuracy of 91.55%. Then, with various feature extraction methods, it was determined that parameters such as MMSE, CDR and SES are important in increasing the accuracy of identification. MR images enhanced with gradient-based filters were classified with 98.63% accuracy using the DenseNet201 model. The Vision Transformer-based model, on the other hand, achieved 85%-92.5% accuracy in classifications such as Alzheimer-Normal and Alzheimer-MCI. In other studies, a hybrid model optimized with Ridge Feature Selection achieved 98.99% accuracy and performed image segmentation with Otsu thresholding method. In addition, the ResNet-based model optimized with SMA achieved 97.48% accuracy and performed particularly well in the MCI class. Finally, a VGGNet-based model optimized with mRMR achieved 98.59% accuracy in four class discriminations, with 100% accuracy, especially in the moderate dementia class. These results show that the proposed methods offer high accuracy and efficiency in Alzheimer's diagnosis and contribute to the literature.