LEVERAGING 3D WAVELET TRANSFORM AND 3D CNN FEATURES FOR EFFICIENT ALZHEIMER’S DISEASE DETECTION

Alzheimer’s Disease (AD) is a progressive neurodegenerative disorder accompanied by the symptoms of gradual decline in socio-cognitive behavioral abilities, which is most commonly seen with aging and these symptoms worsen in advanced stages. The statistics of the World Health Organization (WHO) reveal that globally over 50 million people are experiencing dementia, among which 60–70% cases account for AD. As per the WHO estimates, there is an economic burden of AD $1 trillion annually and it is expected to more than double by 2050. This is a huge burden on the dependent’s family, society, and health care systems. Existing literature has paid attention to computer-aided systems of AD diagnosis to enhance the early detection accuracy through the reduction of cognitive bias. In addition to the existing vast literature on computer-aided Alzheimer’s diagnosis systems, there are several challenges, including security issues, concerns about the quality of available datasets, and biases in algorithms limiting their generalization and diagnostic accuracy. To address this gap, we propose a novel diagnostic framework that integrates wavelet-based feature extraction with a 3D Convolutional Neural Network (3D CNN) architecture, coupled with an ensemble learning strategy. The novelty of the proposed method lies in the integration of wavelet-based feature extraction with a 3D CNN architecture, combined with an ensemble learning strategy. Unlike conventional CNN approaches that may struggle with capturing multi-resolution information in complex neuroimaging data, our model leverages wavelet transforms to efficiently process nonlinear brain patterns across multiple scales. This enables more precise identification of subtle atrophic changes linked to AD. Furthermore, we introduce a majority voting-based ensemble learning framework that combines diagnostic outputs from multiple independently trained models. This ensemble strategy not only enhances diagnostic accuracy but also reduces individual model biases, leading to more robust and generalizable outcomes. Together, these innovations offer a comprehensive and novel pipeline for improving AD diagnosis and tailoring personalized treatment plans. We experimented with the proposed method on two publicly available datasets, ADNI1 and MIRIAD. The [Formula: see text]1-score and sensitivity offered by the proposed model on the ADNI1 dataset are 99.89, 99.85, respectively; on the MIRIAD dataset, they are 90.22 and 95.47, respectively. These results demonstrate the superiority of the proposed model over the state-of-the-art models under comparison.