Deep learning based analysis for Alzheimer's diagnosis

Alzheimer is among the leading causes of death in every country in the world. However, Alzheimer is a difficult disease to diagnose because its symptoms can often be confused with other signs of old age. While there is no single definitive method for diagnosing Alzheimer, it is important to carry out a comprehensive assessment that takes into account multiple factors. In particular, early diagnosis of Alzheimer's disease is essential for both the patient and their relatives to better manage and plan for the disease. In this way, the necessary precautions and measures can be taken. The rapid rise of machine learning, deep learning, and other forms of artificial intelligence is helping to drive major advances in medical image processing. In parallel, research in the field of artificial intelligence is gaining increasing intensity. Deep learning is a machine learning approach that consists of many layers of artificial neurons inspired by human neural networks. Deep learning-based methods have been extensively used in many fields, from translation to text authoring services, from autonomous driving systems to industrial robots, and from healthcare to marketing and sales. In this thesis, new architectures for the early diagnosis of Alzheimer's disease are proposed using different machine learning-based approaches. The proposed architectures utilize convolutional neural network methods for Alzheimer's diagnosis. Brain scan images are classified into three categories with high accuracy: Alzheimer's, mild cognitive impairment, and control groups. The dataset is augmented with various data augmentation methods to train the models. The data is tested with six different models using a novel convolutional neural network architecture with different layers and various transfer learning-based feature extractors. Furthermore, to improve the performance of the transfer learning-based methods, they are connected to a novel neural network-based classifier. In this way, many different methods are tested in this thesis. As a result, a successful Alzheimer's early diagnosis model with high accuracy rates was obtained when our designed classifier was connected to our Xception and VGG19-based pre-trained models.