Abstract
Solving medical imaging data scarcity through semantic image generation hasattracted significant attention in recent years. However, existing methodsprimarily focus on generating whole-organ or large-tissue structures, showinglimited effectiveness for organs with fine-grained structure. Due to stringenttopological consistency, fragile coronary features, and complex 3Dmorphological heterogeneity in cardiac imaging, accurately reconstructingfine-grained anatomical details of the heart remains a great challenge. Toaddress this problem, in this paper, we propose the Fine-grained Cardiac imageSynthesis(FCaS) framework, established on 3D template conditional diffusionmodel. FCaS achieves precise cardiac structure generation using Template-guidedConditional Diffusion Model (TCDM) through bidirectional mechanisms, whichprovides the fine-grained topological structure information of target imagethrough the guidance of template. Meanwhile, we design a deformable MaskGeneration Module (MGM) to mitigate the scarcity of high-quality and diversereference mask in the generation process. Furthermore, to alleviate theconfusion caused by imprecise synthetic images, we propose a Confidence-awareAdaptive Learning (CAL) strategy to facilitate the pre-training of downstreamsegmentation tasks. Specifically, we introduce the Skip-Sampling Variance (SSV)estimation to obtain confidence maps, which are subsequently employed torectify the pre-training on downstream tasks. Experimental results demonstratethat images generated from FCaS achieves state-of-the-art performance intopological consistency and visual quality, which significantly facilitates thedownstream tasks as well. Code will be released in the future.