Virtual and real-world implementation of deep-learning-based image denoising model on projection domain in digital tomosynthesis and cone-beam computed tomography data

Abstract

Reducing the radiation dose will cause severe image noise and artifacts, and degradation of image quality will also affect the accuracy of diagnosis. To find a solution, we comprise a 2D and 3D concatenating convolutional encoder-decoder (CCE-3D) and the structural sensitive loss (SSL), via transfer learning (TL) denoising in the projection domain for low-dose computed tomography (LDCT), radiography, and tomosynthesis. The simulation and real-world practicing results show that many of the figures-of-merit (FOMs) increase in both projections (2-3 times) and CT imaging (1.5-2 times). From the PSNR and structural similarity index of measurement (SSIM), the CCE-3D model is effective in denoising but keeps the shape of the structure. Hence, we have developed a denoising model that can be served as a promising tool to be implemented in the next generation of x-ray radiography, tomosynthesis, and LDCT systems.

Keywords: 3D concatenating convolutional encoder-decoder (CCE-3D); denoising model; low-dose computed tomography (LDCT); structural sensitive loss (SSL); transfer learning (TL).