Latent Diffusion Model for Controllable Lung Nodule Synthesis in CT Scans

A novel controllable latent diffusion model has been developed to generate pulmonary nodules within complete 3D CT volumes, effectively modeling the intensity distributions specific to each nodule. This technique tackles the limited availability of varied annotated datasets for pulmonary nodules by creating synthetic nodules that exhibit realistic texture and attenuation properties across solid, part-solid, and ground-glass subtypes. Current conditional diffusion techniques focus on optimizing spatial reconstruction losses, which often fail to adequately regulate lesion-level intensity distributions, leading to overly smoothed textures. The new method incorporates histogram regularization to ensure precise intensity distributions. This study is available on arXiv with the identifier 2605.30631.