Noise and anatomy-guided diffusion model for realistic CT image synthesis

ICLR 2026 Conference Submission20875 Authors

Keywords: CT Image, Synthesis, Diffusion, Dual-Conditioned

TL;DR: We propose a novel framework for CT image synthesis guided by a noise map and an anatomy mask using a dual conditioning technique for preserving realistic noise-aware anatomical properties.

Abstract: Generative models, specifically Diffusion Models (DMs), have been quite successful in generating high-quality images. However, DMs rely on large-scale training data. In medical imaging, more specifically for computed tomography (CT), these models struggle in accurately reconstructing anatomical structures due to limited training data. This can cause the wrong depiction of organs, which can impact clinical treatment. Some existing models, although guided by anatomical structures, ignore dose-dependent noise, which is critical in real-world scenarios. To tackle this challenge, we propose a novel diffusion model, namely NA-Diff, which is guided by noise from different dose levels and anatomical structures, leveraging a dual conditional diffusion framework. To facilitate large-scale training of DMs on complex structured CT data, we transform natural images emulating realistic CT noises and leverage them for pre-training, followed by fine-tuning on small CT data. Extensive experimental results demonstrate that NA-Diff generates high-fidelity and noise-aware CT images, effectively delineating the organ-of-interest and bridging the gap between synthetic and real CT.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 20875