Advanced CNN architecture integrating machine learning algorithms for precise Alzheimer's disease classification

Abstract

Alzheimer's disease (AD) is a degenerative neurological disorder that affects millions of individuals worldwide and is very difficult to detect and treat in its early stages. This thesis presents a novel architecture for a convolutional neural network (CNN) designed exclusively to classify Alzheimer's disease using functional magnetic resonance imaging (fMRI) data. This work improves the accuracy and reliability of early Alzheimer's identification by using state-of-the-art deep learning techniques to the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. The basis of this research is the ADNI dataset, a vast collection of brain imaging and associated data from people with different degrees of cognitive impairment. The primary objectives are to classify Alzheimer's disease into distinct categories using cognitively normal (CN), early mild cognitive impairment (EMCI), late mild cognitive impairment (LMCI), and Alzheimer's disease (AD) using the recently developed CNN architecture. This study also uses transfer learning techniques to compare the performance of the new CNN with well-known deep learning models like ResNet50 and VGG16, as well as with more conventional machine learning algorithms like XG Boost, k-nearest neighbor (KNN), and Random Forest. The innovative CNN architecture is meticulously designed to maximize classification accuracy. The preprocessing steps involve resizing fMRI images to 109x91 pixels and labeling them accordingly. The network incorporates convolution layers with 3x3 kernels, ReLU activation functions, and 2x2 pooling layers, transforming the images into feature vectors that are subsequently classified. Compared to previous tested models, the innovative CNN architecture performed better, achieving an impressive 99.51% classification accuracy. In terms of comparison analysis, the accuracy of the VGG16 model was 98.24%, whereas the accuracy of the ResNet50 model was 96.05%. The XG Boost classifier, combined with VGG16 for feature extraction, reached an accuracy of 96.93%. The KNN algorithm, also paired with VGG16, exhibited outstanding performance with an accuracy of 98.68%, making it the most effective among the traditional machine learning methods tested. With VGG16 included, the Random Forest classifier produced an accuracy of 94.70%. The outcomes demonstrate how well the suggested CNN architecture performs in comparison to current deep learning and machine learning models in precisely classifying Alzheimer's disease stages. This study demonstrates how sophisticated CNN designs and transfer learning can be used to enhance Alzheimer's disease early detection and diagnosis. The findings suggest that further exploration of alternative deep learning networks, such as convolutional auto encoders, Alex Net, and Google Net, as well as ensemble methods, could enhance model generalization and minimize overfitting. In conclusion, this thesis presents a significant advancement in Alzheimer’s disease classification using fMRI data, providing a robust framework for future research and development in neuroimaging and deep learning applications. The superior performance of the novel CNN architecture demonstrates its potential as a valuable tool for early diagnosis, which is crucial for managing and potentially mitigating the way Alzheimer's disease advances.