Recall-Net: A CNN-based Model for Four-class Classification of Alzheimer’s Disease

Abstract

eep learning, a cutting-edge machine learning technique, has outperformed classical machine learning at detecting detailed structures in complex multi-dimensional data, particularly in the field of computer vision. As recent advances in neuroimaging techniques have created massive multimodal neuroimaging data, the application of deep learning to early diagnosis and automated categorization of AD has recently gotten a lot of interest. It also supports biomedical researchers in the identification of many diseases such as cancer, Alzheimer’s, Malaria, and blood cell detection, among others. Deep learning is a subclass of machine learning techniques for extracting features and applying them to classification, image processing, and other tasks. A thorough Google Scholar search was conducted before and during our research to find deep learning publications on AD published between January 2010 and July 2020. After reading and evaluating, these articles were categorized and summarized according to their used algorithms and neuroimaging techniques. In our research, we have used CNN also known as ConvNet which is one of the most efficient deep learning-based neural networks to classify Alzheimer’s patients from healthy individuals with the help of MRI data. We have collected our sMRI data from ADNI. Our dataset includes a total of 6400 patients who were categorized as non-demented or having mild to severe Alzheimer’s disease. Our deep learning approach automatically distinguishes different stages of AD subjects according to their severity. Our proposed model was designed to assist with accurate classifi cation of four classes e.g. NC, EMCI, MCI and AD. Though classification is very crucial for modeling a prediction model to find out the existence and intensity of the disease, it has always been quite difficult. Separating the distinctive features from the ROI is the most challenging part. Our proposed model has a classification accuracy of 79.77%. The performance of our work was compared to several other existing approaches for multi-class classification.