Enhanced Alzheimer’s Disease Detection Using Transformer-Based GAN and Deep Learning Techniques
TL;DR
An enhanced deep learning framework that integrated a Transformer-based Generative Adversarial Network (TGAN) with an attention-enhanced ResNet-50 architecture for robust multiclass classification of Cognitively Normal, Mild Cognitive Impairment, and Alzheimer’s Disease MRI images is proposed.
Abstract
Early detection of Alzheimer’s Disease (AD) is essential for timely clinical intervention. However, conventional Magnetic Resonance Imaging-based (MRI-based) diagnostic approaches are often limited by class imbalance, insufficient training data, inter-dataset variability, and inadequate feature extraction capability, which reduce the reliability of early-stage AD identification. This study proposed an enhanced deep learning framework that integrated a Transformer-based Generative Adversarial Network (TGAN) with an attention-enhanced ResNet-50 architecture for robust multiclass classification of Cognitively Normal (CN), Mild Cognitive Impairment (MCI), and Alzheimer’s Disease (AD) MRI images. The TGAN model was employed to generate anatomically realistic synthetic MRI scans to address dataset imbalance and improve feature diversity, while the enhanced ResNet-50 utilized residual learning and attention mechanisms to improve the extraction of discriminative neuroanatomical features associated with disease progression. A comprehensive preprocessing pipeline was applied to standardize MRI data obtained from the ADNI and OASIS datasets. We conducted a series of experiments on the ADNI and OASIS databases. The proposed system demonstrated a classification accuracy of 94.8% on ADNI and 93.6% on OASIS, greatly surpassing baseline models such as ordinary ResNet-50, DenseNet-121, and traditional GAN-augmented CNNs. Ablation investigations further verified the significance of the Transformer-based generator and attention-enhanced residual blocks, each yielding substantial performance enhancements.