Alzheimer Disease Classification Using Convolution Neural Networks and Transfer Learning

Abstract

Alzheimer’s Disease (AD) is a neurological disorder which causes brain cells to die, result-ing in memory loss, language diffi culties, and impulsive or erratic behavior. In recent years the number of individuals affected has seen a rapid increase, it is estimated that up to 107 million subjects will be affected by 2050 worldwide. Early diagnosis has become crucial to improve patients care and treatment. AD diagnosis is diffi cult due to the complexity of the brain struc-ture and its pixel intensity similarity especially at its early stage. A comprehensive diagnosis must be led including clinical assessment and medical imaging, which is a process that requires the expertise of professionals including neurologists and radiologists. One of the drawbacks of medical imaging approach is the inability to detect changes in very mild impairment also known as mild cognitive impairment (MCI). Deep learning has inspired a lot of interest in recent years in tackling challenges in a variety of fi elds, including medical imaging and detecting abnormal- ities beyond human capabilities. In this work we explored classifi cation approaches of AD through two different datasets which are respectively MRI dataset and tabular dataset. First, we dealt with the tabular data for bi- nary classifi cation of AD into demented and non demented using classical machine learning algorithms namely Support Vector Machine, K Nearest Neighbor, Decision Tree, Na¨ive Bayes Gaussian, Random Forest and Logistic Regressor. The fi ndings indicate that the models ef- fectively utilize the Clinical Dementia Rating feature for AD classifi cation. Second, we dealtwith MRI dataset for multiclassifi cation of AD into Non Demented, Very Mild Demented,Mild Demented, and Moderate Demented using transfer learning models namely VGG19, Res-Net50, Xception and MobileNet. The VGG19 model gave the best performance with 98.60%testing accuracy where the other models achieved 97.35%, 86.35% and 95.50% respectively.We also proposed a custom CNN model that outperformed the transfer learning models and achieved an accuracy of 99.00%.