Anatomy-guided Latent Diffusion Model for Fine-grained Medical Image Synthetic Augmentation

Sang-Heon Lim, Su Yang, Jiyong Han, SuJeong Kim, DaEl Kim, Yu-Ri Kim, Jun-Min Kim, Jo-Eun Kim, Won-Jin Yi

Keywords: latent diffusion model, latent blending, anatomical synthesis

Abstract: Medical data typically requires expert annotation to produce a reliable quantitative organ analysis, which can be costly and time-consuming. Recently, several deep learning-based synthetic augmentations have been proposed to address the limitations. However, previous success of generative synthetic augmentation methods cannot be guaranteed without additional fine-tuning. To mitigate the dependency on this issue, we propose an anatomy-guided latent diffusion model, which can perform anatomical synthesis in a selectively latent blending manner. We evaluate the proposed approach using a mandibular canal segmentation dataset on panoramic dental radiographs. The segmentation performance was improved by a Dice similarity coefficient of 16.6\% with our proposed synthetic augmentation.

Submission Number: 21