Cycle-Consistent Generative Adversarial Network Based Approach for Denoising CT Scan Images

Abstract

Computed Tomography (CT) plays a pivotal role in detecting and observing various medical conditions. Although low-dose CT scans are frequently chosen to minimize radiation exposure, the image quality is degraded and substantial noise is produced. Important details may be obscured by the added noise. It is challenging for conventional image denoising techniques to strike a compromise between minimizing noise and keeping image details. Hence, minimizing radiation exposure while maintaining image quality makes denoising low-dose CT images a critical challenge in medical imaging.

This study investigates the application of CycleGAN, a deep learning-based model, for this purpose. CycleGAN model makes use of a U-net based generator and PatchGAN based discriminator. A conventional UNet model and a Gaussian filter were compared with the CycleGAN model. Structural Similarity Index (SSIM) and peak signal-to-noise ratio (PSNR) measures were utilized to assess the models’ functioning in terms of quality of image and structural similarity. The CycleGAN model performed better than the UNet and Gaussian filters, according to the results, striking an improved harmony between detail preservation and noise reduction. The results indicate that CycleGAN is a promising low-dose CT image denoising technique that can be used to improve diagnostic accuracy while exposing users to as little radiation as possible.