Dual-WGAN Ensemble Model for Alzheimer’s Dataset Augmentation with Minority Class Boosting

Deep learning models have become very efficient and robust for several computer vision applications. However, to harness the benefits of state-of-art deep networks in the realm of disease detection and prediction, it is crucial that high-quality datasets be made available for the models to train on. This work recognizes the lack of training data (both in terms of quality and quantity of images) for using such networks for the detection of Alzheimer’s Disease. To address this issue, a Wasserstein Generative Adversarial Network (WGAN) is proposed to generate synthetic images for augmentation of an existing Alzheimer brain image dataset. The proposed approach is successful in generating high-quality images for inclusion in the Alzheimer image dataset, potentially making the dataset more suitable for training high-end models. This paper presents a two-fold contribution: (i) a WGAN is first developed for augmenting the non-dominant class (i.e. Moderate Demented) of the Alzheimer image dataset to bring the sample count (for that class) at par with the other classes, and (ii) another lightweight WGAN is used to augment the entire dataset for increasing the sample counts for all classes.