关键词: 潜在扩散模型, 频域损失, 支持带约束, 潜空间混合

Abstract:

Brain magnetic resonance imaging (MRI) is widely used in the medical field; however, in small-sample scenarios, challenges such as insufficient training data, the limited distribution expansion capability of traditional data augmentation methods, and the structural distortion and sample memorization issues of diffusion models remain significant. To address these problems, this paper proposes a medical image generation diffusion model under cross-sample latent structure reorganization constraints. Built upon a latent diffusion framework, the proposed method introduces a cross-sample block-wise latent feature mixing mechanism in latent space to improve the distribution coverage and structural diversity of generated samples. In addition, a spectral consistency constraint is incorporated to enhance the preservation of high-frequency textures and edge details, while a support-band constraint is introduced to prevent generated samples from deviating from the real medical image manifold, thereby improving the structural realism and medical plausibility of generated images. Experimental results demonstrate that the proposed method can generate medical images with high realism, diversity, and structural consistency across multiple medical imaging datasets, while effectively improving the performance of downstream brain tumor detection.

Key words: Potential diffusion models, frequency domain loss, support band constrained, latent space mixing