Enhanced medical image generation through advanced latent space diffusion

Collecting high-quality medical image data for machine learning applications remains a significant challenge due to data scarcity, privacy concerns, and high annotation costs. To address these issues, vision generative models, particularly Latent Diffusion Models (LDMs), have emerged as state-of-the-art solutions that reduce computational demands while maintaining superior performance in data generation tasks. In this study, we propose an enhanced LDM-based approach that integrates separable self-attention mechanisms within the diffusion process, positioned after residual blocks, to improve the capture of detailed features and maintain spatial consistency. This modification reduces memory usage by 82.94% and decreases the Fréchet Inception Distance (FID) by 25.01% compared to traditional self-attention models, all while preserving image quality. Our method addresses critical challenges such as data scarcity and computational efficiency in medical imaging by combining Variational Autoencoders (VAEs) for latent space mapping with U-Net for noise prediction. Evaluations on five datasets — PneumoniaMNIST, BloodMNIST, ChestMNIST, Dental4k, and HandMNIST — demonstrate significant improvements in computational efficiency, memory usage, and the quality of generated images, showcasing the potential of our approach for scalable and effective medical image synthesis.