Diffusion-based image translation model from low-dose chest CT to calcium scoring CT with random point sampling

This study proposes a diffusion model-based domain adaptation method to improve coronary artery calcium (CAC) scoring using low-dose chest CT (LDCT). Due to high noise and low resolution, LDCT is less accurate than calcium scoring CT (CSCT). To address this, the authors developed a conditional diffusion model based on the denoising diffusion implicit model (DDIM), introducing two improved sampling methods: random pointing and intermediate sampling. This method effectively reduces noise while preserving calcium structures and reduces the number of sampling iterations from 1000 to 10, increasing clinical applicability. Aview, a cardiac detection model by Coreline Soft, was used to crop LDCT and CSCT data for training. The proposed model outperformed existing image-to-image translation methods, such as CycleGAN and CUT, in major metrics including PSNR and SSIM.