An NCA-based Hybrid CNN Model for Classification of Alzheimer’s Disease on Grad-CAM-enhanced Brain MRI Images

Abstract

Alzheimer’s, one of the most prevalent varieties of dementia, is a fatal neurological disease for which there is presently no known cure. Early diagnosis of such diseases and classification with computer-aided systems are of great importance in determining the most appropriate treatment. Imaging the soft tissue of the brain with Magnetic Resonance Imaging (MRI) and revealing specific findings is the most effective method of Alzheimer’s diagnosis. A few recent studies using Deep Learning (DL) to diagnose Alzheimer’s Disease (AD) with brain MRI scans have shown promising results. However, the fundamental issue with DL architectures like CNN is the amount of training data that is required. In this study, a hybrid CNN method based on Neighborhood Component Analysis (NCA) is proposed, which aims to classify AD over brain MRI with Machine Learning (ML) algorithms. According to the classification results, DenseNet201, EfficientNet-B0, and AlexNet pre-trained CNN architectures, which are 3 architectures that give the best results as feature extractors, were used as hybrids among 10 different DL architectures. By means of these CNN architectures, the features trained on the dataset and the features obtained by Gradient-weighted Class Activation Mapping (Grad-CAM) are concatenated. The NCA method has been used to optimize all concatenated features. After the stage, the optimized features have been classified with KNN, Ensemble, and SVM algorithms. The proposed hybrid model achieved 99.83% accuracy, 99.88% sensitivity, 99.92% specificity, 99.83% precision, 99.85% F1-measure, and 99.78% Matthews Correlation Coefficient (MCC) results using the Ensemble classifier for the 4-class classification of AD.