Low‐dose computed tomography scheme incorporating residual learning‐based denoising with iterative reconstruction

Abstract

Low-dose computed tomography has been highly desirable because of the health concern about excessive radiation dose, but also challenging due to insufficient or noisy projection data. Compared with post-processing methods by directly denoising filtered backprojection images, iterative reconstruction achieves excellent performance but consumes a large number of iterations. In this Letter, a two-stage method is proposed by incorporating residual learningbased denoising with iterative reconstruction. First, an intermediate image is reconstructed by compressed sensing iterative reconstruction. Then, the image is denoised by a deep neural network. Specially, a network performing two-level residual learning is designed to strengthen denoising effect. Experimental results show that the proposed method outperforms iterative reconstruction with better numeric results and comparable visual performance while consuming fewer iterations.

An improved non-local means algorithm for CT image denoising

“…As an important method of medical diagnosis and treatment, computed tomography (CT) is of great value in disease screening. And it has been successfully used in clinical diagnosis [1] . However, in the process of collecting CT data, it is inevitable to be interfered by noise, which leads to the blurred image or the loss of detail structure [2] .…”