Early and accurate detection of Alzheimer’s disease (AD), particularly at the very mild and mild cognitive impairment stages, remains a significant clinical challenge due to subtle anatomical changes and overlapping imaging features between classes. This study evaluates the effectiveness of synthetic MRI image augmentation using Deep Convolutional Generative Adversarial Networks (DCGANs) to improve the classification performance of deep learning models for early-stage AD detection. The dataset consists of Magnetic Resonance Imaging (MRI) scans categorized into “No Dementia,” “Very Mild Dementia,” “Mild Dementia,” and “Moderate Dementia” classes, sourced from a publicly available repository. A baseline Convolutional Neural Network (CNN) was initially trained on traditionally augmented images and evaluated on a validation set. To address class imbalance and improve sensitivity for the “Very Mild Dementia” class, a DCGAN was trained on this subset to generate synthetic MRI images. Generator models were checkpointed based on their Frechet Inception Distance (FID) scores, and 500 synthetic images from each selected generator were incorporated into the training set. Comparative analysis revealed that while the baseline CNN achieved a validation accuracy of 98% and a precision of 0.97 for the “Very Mild Dementia” class, the augmented model utilizing synthetic images from the Epoch 72 generator improved performance to 99% accuracy with a macro F1-score of 1.00. Statistical significance was confirmed via McNemar’s test ( p < 0.05), highlighting the potential of GAN-based augmentation for enhancing early AD classification. These findings underscore the importance of strategic checkpoint selection in GAN-based data augmentation to ensure clinically meaningful performance improvements.
Shetty et al. (Mon,) studied this question.