Noise-conscious explicit weighting network for robust low-dose CT imaging

Abstract

Supervised deep learning (DL) methods have been widely developed to remove noise-induced artifacts and promote image quality in the low-dose CT imaging task via good mapping capabilities. These supervised DL methods are usually trained based on a large amount of low- and normal-dose sinogram/image pairs and the reconstruction performance of such supervised DL methods heavily depends on the quality of reference training images. In the CT imaging, it is challenging to collect lots of high quality reference training images in practice due to the risk of high radiation dose to patients. Moreover, the medium quality or even low quality reference training images (i.e., the low quality labeled data with some noise-induced artifacts) might be collected for the supervised DL networks training, which would degrade the reconstruction performance of the network. To address this issue, in this work, we propose an effective noise-conscious explicit weighting network (NEW-Net) for low-dose CT imaging wherein the CT images with noise-induced artifacts are treated as labeled data in the network training. Specifically, the proposed NEW-Net consists of two sub-networks, i.e., noise estimation sub- network, and noise-conscious weighting sub-network. The noise estimation sub-network produces the noise map from the low-quality training data to estimate the noise-conscious weights, which determines the contribution of the label data, i.e., small weights go along with the low quality label data with severe nosie-induced artifacts, and large weights go along with high quality label data with a few noise-induced artifacts. Then the estimated weights are used to condition the training data to train the noise-conscious weighting sub-network to eliminate the effects of low quality label data and promote the reconstruction performance and stability of the proposed NEW-Net method. The Mayo clinic data are utilized to validate and evaluate the reconstruction performance of the proposed NEW-Net method. And the experimental results demonstrate that the proposed NEW-Net method outperforms the other competing methods in the case of low quality training data, in terms of noise-induced artifacts and structure detail preservation both qualitatively and quantitatively.