Anatomy-compliant medical image synthesis by latent diffusion models

Nuno Miguel Ferreira Capitão, Yidong Zhao, Yi Zhang, Nathan Geerts, João Viana Lopes, Qian Tao

Keywords: Latent diffusion models (LDM), Medical image synthesis, Edge

Abstract: Data scarcity presents a significant challenge for achieving optimal model performance in medical imaging, due to the limited availability of high-quality data. One potential solution to address this issue is to synthesize medical images using powerful generative models with conditioning prior. However, obtaining full anatomical annotations of all organs for anatomical conditioning is impractical, resulting in synthetic images with incoherent or hallucinated anatomy. In this paper, we propose an innovative medical image generation method based on state-of-the-art latent diffusion models (LDM). To tackle the anatomy compliance challenge, we leverage both the anatomical mask, which is specific to the organ of interest, and the edge information, which is general and easy to compute in the full field of view (FOV), as dual conditioning. Our method does not require extra annotations to achieve anatomy compliance. Our method was evaluated on the ACDC dataset and compared with GAN baselines. Results demonstrate that incorporating edge-based conditioning strongly complements image semantics, leading to high-quality, anatomy-compliant medical image generation.

Submission Number: 143