An evaluation of data balancing techniques for improved performance in Alzheimer's disease diagnosis

Early diagnosis of Alzheimer's disease is of critical importance for ensuring the effectiveness of the treatment process. However, class imbalance in medical imaging data can negatively affect the performance of machine learning and deep learning–based classification models. This thesis investigates the class imbalance problem in Alzheimer's disease by comparatively analyzing three different synthetic data generation techniques SMOTE, ADASYN, and Conditional GAN (CGAN) applied to Magnetic Resonance Imaging (MRI) data. The study utilizes an Alzheimer's MRI dataset obtained from the Kaggle platform, consisting of four classes: Non-Demented, Very Mild Demented, Mild Demented, and Moderate Demented. After preprocessing, a 2D CNN-based classification model with an identical architecture was trained separately for each balancing method, and model performance was evaluated using metrics such as accuracy, precision, recall, and F1-score. The results demonstrated that the applied balancing techniques had significant effects on classification performance from both numerical and visual perspectives. This thesis aims to contribute to addressing the class imbalance problem and to serve as a guide for deep learning–based classification studies in the field of medical imaging.