Improved VAE-GAN via mixture of Gaussians applied to brain tumor MRI classification

Generative models have greatly advanced artificial intelligence by making it possible to create realistic data. However, these models often face mode collapse. This limitation prevents them from capturing the full diversity of the data, especially when trained on small datasets. This weakness is particularly noticeable in models like VAE-GAN, where using a simple (unimodal) Gaussian distribution as the sampling prior for generation is not enough to represent complex and varied data. This reduces the diversity and quality of the generated samples. To tackle this issue, we propose a novel generative framework called M-VAEGAN, which replaces the sampling prior for the GAN generator with a mixture of Gaussians to enable a more expressive and flexible latent space representation. M-VAEGAN incorporates regularization and jointly optimizes the mixture parameters alongside the VAE-GAN generator. This enhances its training stability and allows it to achieve better coverage of diverse data modes. Extensive experiments on CIFAR-10, toy data, freehand sketch and CelebA datasets demonstrate that our approach improves generative diversity compared to baseline GAN models. Furthermore, we apply M-VAEGAN in a real world application for brain tumor MRI augmentation. The results show that M-VAEGAN produces high fidelity synthetic samples that effectively balance class distributions, resulting in improved tumor classification accuracy over other augmentation methods. • M-VAEGAN integrates a Mixture of Gaussians prior to improve generative diversity. • Joint training of the parameters of the MoG mitigates mode collapse. • Generates realistic and diverse brain tumor MRI samples to balance datasets. • Improves classification accuracy on imbalanced medical imaging data. • Demonstrates superior performance over conventional generative approaches.