Enhancement of Alzheimer’s Disease Detection Technique by Fusing Features of Deep Learning and Handcrafted Methods on sMRI Images

Automated detection of Alzheimer disease (AD) is evolving rapidly with the advent of artificial intelligence (AI). Machine learning and deep learning models. Detection of AD at an early stage provides advance planning of treatment to the patients that arrest progression of the disease. The aim of this paper is to improve performance of diagnosing AD by fusing features extracted mainly from two sources, such as deep and nondeep frameworks applied on the sMRI images. This paper introduces two customized convolutional neural networks (CNN) (baseline model and hybrid model) and extracted two different sets of deep features. As an alternative source of features, a set of handcrafted features are extracted from the sMRI images by applying the stack assembling of Gray Level Co-occurrence Matrix (GLCM) and Gabor algorithms. In order to select nonredundant important features, we apply cosine similarity metric between the handcrafted and deep features. Next, a Feature Pyramid Network-based feature fusion technique namely-FusionNet is also proposed to fuse the deep features and the handcrafted features in order to further improve the classification accuracy. For comparing the performance of AD diagnosis, fused feature sets without feature selection and with feature selection are considered. It has been observed that the fused features without feature selection achieves the best classification accuracy implying that for AD diagnosis all the features are important and therefore distinct. All the experiments have been applied on a 5-class dataset obtained from ADNI and classified by Support Vector Machine classifier.