Early prediction of Alzheimer's disease using convolutional neural network

Abstract

Neuroimaging can be a prospective instrument for the diagnosis of Mild Cognitive Impairment (MCI) along with its more severe stage, Alzheimer's disease (AD). High- dimensional classi cation methods have been commonly used to explore Magnetic Resonance Imaging (MRI) for automatic classi cation of neurodegenerative diseases like AD and MCI. Early AD or MCI can be diagnosed through proper examination of several brain biomarkers such as Cerebrospinal Fluid (CSF), Media Temporal Lobe atrophy (MTL) and so on. Abnormal concentrations of the mentioned biomarkers on MRI images can be a potential sign of AD or MCI. In the recent times, several high- dimensional classi cation techniques have been suggested to discriminate between AD and MCI on the basis of T1-weighted MRI of patients. These techniques have been implemented mostly from scratch, making it really di cult to achieve any meaningful result within a short span of time. Therefore, classi cation of AD is usually a very daunting and time consuming task. In our study, we trained high dimensional Deep Neural Network (DNN) models with transfer learning in order to achieve meaningful results very quickly. We have used three di erent DNN models for our study: VGG19, Inception v3 and ResNet50 to classify between AD, MCI and Cognitively Normal (CN) patients. Firstly, we implemented some pre-processing steps on the images and divided them into training, testing and validation sets. Secondly, we initialized these DNN models with the weights from pre-existing models trained on the imagenet dataset. Finally, we trained and evaluated all the DNN models. After relatively short amount of trainings (15 epochs), we achieved an approximate of 90% accuracy with VGG19, 85% accuracy with Inception v3 and 70% with ResNet50. Thus, we achieved excellent classi cation accuracy in a very short time with our research.