Alzheimer's disease (AD) is a chronic neurodegenerative disorder affecting memory, thinking, and behavior. Deep learning models, particularly CNNs, have shown promise in detecting AD at initial stages using the brain's magnetic resonance images (MRI). In this study, a CNN model called ADNet, trained using the OASIS-1 dataset, was proposed. The experimental approaches for evaluating the performance of ADNet are as follows: First, three different datasets were prepared using slices taken from the first quarter, middle, and third quarter of the sagittal plane from each MRI, to determine the most informative slice among the 128 slices. Each dataset was split into 80% training and 20% testing. It was found that the first quarter slice showed the best performance. The potential use of the obtained model as a transfer learning model was also examined. For this, a low-performance model was retrained using ADNet as a transfer learning model, and significant improvements in the results were observed. At last, the model’s robustness was evaluated in a more detailed evaluation, using 5-fold cross-validation repeated three times, resulting in a mean accuracy of 97.05%. As a result, ADNet can be used for Alzheimer's screening in clinical settings and could enable patients to receive earlier treatment.
Saracoğlu et al. (Tue,) studied this question.