Ultra-low Dose CT Image Denoising based on Conditional Denoising Diffusion Probabilistic model

Abstract

Due to repeated examinations of lung nodules by Standard Dose Computed Tomography (SDCT), patients suffer from an increased risk of further cancer deterioration caused by the accumulated X-ray dose. Although radiologist have attempted using Ultra-low dose CT images instead of SDCT for diagnosis, the reduction of CT dose decreases the final reconstructed image quality and seriously hinders diagnosis. To compensate for the reduced image quality, we presents a novel noise reduction approach, conditional Denoising Diffusion Probabilistic Model (c-DDPM), by exploiting the advantages of Diffusion Probabilistic Models (DDPM). c-DDPM applies a 2.5D feature fusion strategy to account for CT spatial details, and constrains the denoising procession, by combining the loss function l2 and lssim. We evaluate c-DDPM and a state-of-the-art method CycleGAN, the commercial IMR method and iDose on an actual patients dataset with a total of 170 patients. Objective assessment shows that c-DDPM can suppress the isolated artifacts and generate more compelling ULDCT images with PSNR (35.19±0.73) and SSIM (0.85±0.03). The subjective evaluation performed by radiologists also demonstrates that our approach can effectively improve perceptual image quality, achieving an overall image quality score of 4/5 or above in 88.4% of cases and an image noise score of 4/5 or above in 100% of the cases. Finally, we provides comprehensive empirical evidence showing that in the lung nodule detection task, ULDCT images denoised through c-DDPM my be detected 11% more valid nodules than of CycleGAN.