Comparable Study of Pre-trained Model on Alzheimer Disease Classification

Abstract

The Alzheimer's disease (AD) is a type of dementia that affects millions of people worldwide every year and the occurrence will continue to be on the increase. The move to diagnose people suffering from AD at an earlier stage has been a daunting problem in mental health. In recent years, the advancement of deep learning in the likes of convolutional neural networks (CNN) has made a great effort towards an early detection of AD using magnetic resonance imaging (MRI) data. However, due to the need for highly discriminative features from MR images, it is still challenging to accurately use CNNs by training from scratch for early detection of AD. This paper aims to improve the early detection of Alzheimer's disease using deep learning for neuroimaging data. We have utilized the SqueezeNet, ResNet18, AlexNet, Vgg11, DenseNet, and InceptionV3 pre-trained models to automatically classify MR images. To validate our model, we experimented with the MR images obtained from the Open Access Series of Imaging Studies (OASIS) database. The average classification accuracy derived by SqueezeNet model for training and testing was 99.38% and 82.53% for binary class and multiclass, respectively.