Alzheimer's disease classification using 3D conditional progressive GAN- and LDA-based data selection

Abstract

Alzheimer's disease is a kind of neurological disorder that directly impacts the memory of a patient. Structural magnetic resonance imaging (sMRI) is an effective representation for the diagnosis of neurodegenerative diseases. Deep learning strategies, such as convolutional neural networks (CNNs), require an enormous amount of data to generalize the target disease. Given the restrictions on collecting data, augmentation methods are important tools for increasing the number of samples available for training a CNN. Recently, generative adversarial networks (GANs) have been employed to generate synthetic medical data such as sMRI. In this paper, we propose a conditional progressive GAN (cProGAN) for data augmentation. The proposed cProGAN utilizes additive noise, which is regulated by the feedback from the discriminator that is trained by labeled data. The synthetic samples generated by using cProGAN go through a sample selection process regulated by the distributions of the original data mapped into the linear discriminator analysis (LDA) space. Three-class labeled data are mapped into LDA space where each class is modeled within an elliptic confidence subspace. Generated synthetic data that falls into these class subspaces are selected as the synthetic data to be used for training the CNN. This strategy helps select the most relevant samples with the desired class. Evidently, based on the experimental results, the suggested cProGAN creates synthetic data with higher quality than other state-of-the-art approaches. Furthermore, class-specific LDA subspace post-processing helps the selection of class-separated augmented data for improved classification performance.