EARLY DIAGNOSIS OF ALZHEIMER’S DISEASE: EFFECTS OF REGISTRATION, BRAIN EXTRACTION AND AUGMENTATION OF 3D MRI SCANS COMBINED WITH NEURAL NETWORK PARAMETER TUNING

Alzheimer’s Disease (AD), considered as the most widespread form of dementia, is a neurological condition that results in a progressive brain atrophy and mental deterioration. Even though an effective treatment for AD still not exists, an early detection of this pathology may help in slowing down the disease progression. Among all available neuropsychological data, T1-weighted structural magnetic resonance images, which stress the difference between white and grey matter, appear to be the best choices to recognize the disease. In recent years, an increased interest in AD detection from magnetic resonance images using artificial intelligence algorithms have emerged. Specifically, many deep learning approaches, which integrates the feature step in the learning one, have been proposed, most of them involving 3D convolutional neural networks. This work is intended to investigate the effects of registration, brain extraction and data augmentation of magnetic resonance scans, together with the parameter tuning of a Convolutional Long-Short Term Memory (ConvLSTM)-based neural architecture, on the model performance for AD classification. To this aim, 275 scans were selected from the OASIS-3 dataset (145 AD and 130 cognitively-healthy patients). Pre-processing consisted in linear affine registration with three different templates, brain extraction and data augmentation (horizontal flipping and rotation). Nine experiments were conducted: each experiment, expect for the first, started from the previous experiment and added something new. First, the performance of the proposed framework was quantified on the raw data. Then, the impact of the registration step using three different templates was evaluated, leading to the identification of the best one. Brain extraction, data augmentation and, finally, parameter tuning were gradually added till the last experiment. Results show how the model performance tend to increase step by step until the parameter tuning, which lead to an accuracy of 85%, a sensitivity of 88% and an area under the curve of 92%. In conclusion, the work hereby presented demonstrates the crucial importance of pre-processing steps such as registration and brain extraction, data augmentation and parameter tuning in enhancing the AD classification performance of the ConvLSTM-based framework.