Advanced Deep Learning Frameworks for Early Alzheimer’s Disease Prediction: A Comparative Analysis of Hybrid Architectures and Transfer Learning on MRI Data

Abstract

Alzheimer disease (AD) is a progressive neurodegenerative disease and the most prevalent type of dementia, which is the incurable impairment of the memory and cognitive functions. There is an imminent need to diagnose dementia early and accurately as the prevalence of this condition is likely to hit 152 million around the world by 2050. Conventional diagnostic procedures that are based on hand reading of Magnetic Resonance Imaging (MRI) are also timely and subject to observer bias. The current paper explores the effectiveness of Deep Learning (DL) algorithms in the classification of AD. Our results on investigating the performance of Transfer Learning with pre-trained Convolutional Neural Networks (CNNs) like AlexNet, ResNet, and EfficientNet, and novel hybrid architectures combining CNNs with Vision Transformers (e.g., Swin Transformer), evaluate the capability of identifying both local anatomical and global contextual dependencies, and we assess the black box nature of DL through Explainable AI methods such as Grad-CAM to promote clinical interpretability. The experimental findings on the ADNI and OASIS data sets show that the hybrid models and strict data augmentation (e.g., DCGAN) can significantly enhance the quality of diagnosis and the ability to overcome the class imbalance.