Alzheimer disease classification using MRI images based on transfer learning

Abstract

Alzheimer’s disease (AD) affects cognitive skills and progresses as a result of neurodegenerative changes in the brain region. Until clinical symptoms usually begin, early detection of AD is critical to ensure timely treatment. Computer-assisted diagnosis (CAD) is an essential aid in early AD diagnostics allowing for simpler and more efficient therapies. Machine learning in recent times has really been used to detect AD based on data from neuroimaging including MRI. Deep Learning (DL) models, on the other hand, permit a computer to use dataset in raw form as an input, helping it to identify extremely complex features from the training dataset without the intervention of humans. However, such models are often limited by the need for a vast amount of training imaging data as well as the appropriate optimization of the layout. This paper address these problems adopting transfer training wherein the state-of-the-art VGG 19 and ResNet50 networks have been deployed. These architectures are already learned on large ImageNet baseline collection containing pre-trained weights. The systems are therefore refined with the layer-specific tuning, which only a predefined group of layers on neuroimaging dataset particularly MRI. The performance of these two architectures is compared for two binary classification problems AD vs. Mild cognitive impairment (MCI) and Normal Control (NC) vs. MCI as these two are crucial as compared to AD vs. NC. Our methodology is validated on the dataset collected from the AD disease Neuroimaging Initiative (ADNI) database, which includes 20 AD, 20 MCI, and 20 NC cohorts for baseline T1-weighted structural MRI data. However, VGG 19 architecture outperformed ResNet50 for various batch sizes. Furthermore, VGG 19 architecture with batch size 16 is more accurate, with an accuracy rate of 93.89% for AD vs. NC, as well as 92.89% for the NC vs. MCI and 89.4% for AD vs. MCI in comparison to batch size 32.