Cycle generative adversarial network-assisted alzheimer's disease diagnosis using magnetic resonance images

Alzheimer's disease is a neurodegenerative disease affecting millions of people worldwide and causing loss of cognitivie functions primarily in older individuals. Magnetic Resonance Imaging (MRI) and Positron Emission Tomography facilitate disease monitoring by revealing changes in different brain regions. Early diagnosis is critical for treatment success and improving patients' quality of life. MRI analysis is widely used for early Alzheimer's diagnosis, enabling staging and classification of the disease. However, current diagnostic methods often rely on limited datasets, necessitating bigger datasets to enhance diagnostic accuracy. In this study, the unsupervised transition from image to image between two classes of MR images, namely those with no Alzheimer's symptoms and those with very mild Alzheimer's symptoms, was investigated. Bidirectional synthetic data was generated by using Cycle Generative Adverserial Networks (CycleGAN). Synthetic data augmentation was performed by converting MRI images without Alzheimer's symptoms into images with very mild Alzheimer's symptoms and performing the inverse transformation, and 100 MRI images of each class were obtained. The performances of transfer learning-based binary classification on the original dataset, the dataset extended with CycleGAN, and the dataset produced only with CycleGAN were demonstrated. Performance increase was not observed in the CycleGAN extended dataset compared to the original dataset. In this thesis, the advantages of using generative adversarial networks in the field of computer-aided Alzheimer's diagnosis were investigated.