A SEMANTIC EMBEDDING VARIATIONAL AUTOENCODER GENERATIVE ADVERSARIAL NETWORK MODEL FOR ALZHEIMER’S DISEASE MRI IMAGE CLASSIFICATION

Alzheimer’s disease (AD) is a neurodegenerative disorder characterized by irreversible damage to memory-related cells. Early detection of AD through magnetic resonance imaging (MRI) is a challenging task. To this end, this paper proposes an MRI image classification model for AD based on semantic embedding variational autoencoder generative adversarial network (SVAE-GAN). The SVAE-GAN model is an enhancement of traditional VAE that incorporates adversarial learning to improve both the reconstruction and generation capabilities of the model. This integration allows the model not only to generate high-quality images but also to ensure that the generated images are realistic and diverse. During the training of the autoencoder, the image label information is utilized in a supervised learning manner to establish a classifier, enabling the constructed latent encoding space to extract discriminative potential features that are relevant for classification. Subsequently, during the training of the GAN, the data distribution learned by the generator in the latent space approximates latent variable space by integrating a generative adversarial mechanism, to construct high-quality image samples. The gradient normalization constraint is added to the discriminator, which is further used to improve the network stability. Finally, the generated pseudo data, together with real MRI image data, is used for training, achieving early detection of AD based on MRI images. The experiment validates that the SVAE-GAN model effectively improves performance compared to the current mainstream image generation and classification methods.