A model to improve the quality of low-dose CT scan images

Abstract

Computed Tomography (CT) scans are used during medical imaging diagnosis as they provide detailed cross-sectional images of the human body by making use of X-rays. X-ray radiation as part of medical diagnosis poses health risks to patients leading experts to opt for low doses of radiation when possible. In accordance with European Directives, ionising radiation doses for medical purposes are to be kept as low as reasonably achievable (ALARA). While reduced radiation is beneficial from a health perspective, this impacts the quality of the images as the noise in the images increases, reducing the radiologist’s confidence in diagnosis. Various low-dose CT (LDCT) image denoising strategies available in the literature attempt to solve this conflict. However, current models face problems like over-smoothed results and lose detailed information. Consequently, the quality of LDCT images after denoising is still an important problem. The models presented in this work use deep learning techniques that are modified and trained for this problem. The results show that the best model in terms of image quality achieved a peak signal to noise ratio (PSNR) of 19.5 dB, a structural similarity index measure (SSIM) of 0.7153 and a root mean square error (RMSE) of 43.34. It performed the required operations in an average time of 4843.80s. Furthermore, tests at different dose levels were done to test the robustness of the best performing models.