A Methodology to Train a Convolutional Neural Network-Based Low-Dose CT Denoiser With an Accurate Image Domain Noise Insertion Technique

To mitigate the public health risks associated with the increasing utilization of computed tomography (CT), it is desirable to implement a low-dose scanning protocol. However, low-dose CT produces poor image quality due to the increased quantum noise, requiring an effective image denoising method for its clinical use. Recently, convolutional neural network (CNN)-based methods showed state-of-the-art performance in CT image denoising. However, a large amount of paired training data is essential for their high performance, which is generally not available in medical imaging fields. To alleviate this problem, we propose a new framework to train a CNN-based denoiser without real paired CT images or access to the proprietary information such as raw scan data and a reconstruction kernel. Specifically, we estimate a reconstruction kernel from normal-dose CT images and synthesize paired CT images using an image domain noise insertion technique. To validate the proposed method, we used extended cardiac-torso phantoms and the 2016 Low-Dose CT Grand Challenge datasets. The denoising performance of the CNN-based denoiser trained with the proposed method was compared with the conventional approach that generated paired training data by adding noise to raw scan data and reconstructing images with the known reconstruction kernel. To quantitatively evaluate the image quality, we used the peak signal-to-noise ratio, structural similarity index measure, modulation transfer function, and noise power spectrum. The extensive qualitative and quantitative results suggest that the proposed method enables CNN-based denoisers to be trained without access to raw scan data and a reconstruction kernel while maintaining their denoising performance.

Abstract licence: CC BY-NC-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.