Diagnosis of Alzheimer’s Disease with Deep Neural Networks

Alzheimer’s disease is the most common type of dementia and, although there is no cure for it, its early diagnosis is essential for effective treatment to slow the progression of symptoms. Therefore, in recent years, there has been great interest in studying and developing automatic diagnostic systems that use medical examinations performed by patients as a data source. This paper addresses the use of deep learning models to diagnose Alzheimer’s disease. It was intended to evaluate the performance of existing deep neural networks as realistically as possible and to propose an architecture with good performance that can be used in a computer-assisted diagnostic system. The ability of deep learning to find hidden patterns in medical images makes it possible to reduce human diagnostic error and enables a much more accurate diagnosis. Based on a single image per patient, the proposed neural network was expected to be able to make a diagnosis of Alzheimer’s disease. The work was carried out with brain images acquired with the Structural Magnetic Resonance Imaging technique, belonging to the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset. It was concluded that convolutional models, trained for the task of classifying magnetic resonance images, are a viable option to diagnose Alzheimer’s disease. In most cases, 3D models perform better than the corresponding 2D models. Among Alzheimer’s disease, mild cognitive impairment, and normal control, mild cognitive impairment cases are the most difficult to identify by the models. The SeResNext50 3D model was the most successful in identifying mild cognitive impairment, obtaining an AUC of 0.70. Training models with images resulting from the application of a higher isotropic resolution did not produce better results than training models with a lower spatial resolution. The data augmentation techniques applied did not improve the results, indicating that the generated images were highly correlated with the original ones. The ADNI data set is not very large, and the decision to rule out samples of the same patient, collected at very close times, led the dataset to become even smaller, allowing only results similar to those reported in the related bibliography.