Enhancing Quality of Low-Dose CT Scans Via Generative Diffusion Models

Abstract

The enduring challenge in computed tomography (CT) imaging is mitigating the radiation risks associated with high-dose protocols while maintaining image quality. This research introduces an innovative approach that diverges significantly from conventional methodologies, using Generative Diffusion Models (GDM) to enhance the quality of low-dose CT scans to that of high-dose scans. This advancement is particularly pivotal as it addresses the crucial balance between minimizing radiation risk and preserving diagnostic integrity. At the heart of our approach is a distinctive application of a Convolutional Neural Network (CNN) designed not to filter noise but to meticulously identify and segregate intrinsic noise features within paired high and low-dose CT images. This method stands in contrast to traditional techniques that often rely on generic random noise models, lacking specificity to actual imaging conditions. By accurately modeling the unique noise profile of low-dose scans, we enable our GDM to undertake a reverse diffusion process, effectively reducing noise and enhancing image clarity to equal high-dose standards. The significance of transitioning from low-dose to high-dose imaging quality without additional radiation is offering a path to safer imaging protocols that do not compromise quality. We present preliminary findings substantiated by both PSNR and SSIM metrics, demonstrating improvement in image quality through our method. In addition to delineating our approach, this research draws comparisons with existing methods, particularly focusing on PALLETE, a known algorithm in the field. Our comparative analysis illustrates the superiority of our model in terms of image quality, showcasing our method’s potential for enhancement in radiological imaging.