DPDudoNet: Deep-Prior Based Dual-Domain Network for Low-Dose Computed Tomography Reconstruction

Low-dose computed tomography (LDCT) reconstruction has been an active research field for years. Although deep learning (DL)-based methods have achieved incredible success in this field, most of the existing DL-based reconstruction models lack interpretability and generalizability. In this paper, we propose a novel deep prior-based dual-domain network (DPDudoNet) by unrolling the model-based algorithm using iteratively-cascaded DenseNet and deconvolutional network. The proposed model embeds the intrinsic imaging model constraints, inherited from the foundational model-based algorithm, to tackle the issue of lacking interpretability. Besides, it contains fewer learnable parameters, compared to many representative networks, thus leading to simpler decision boundary and better generalizability. Moreover, a random initialization of the network based on Gaussian distribution is introduced as a deep prior for the LDCT reconstruction. The proposed model integrates the deep prior into both the image and sinogram domains via a dual-domain update scheme. Experimental results on the public AAPM LDCT dataset show that our proposed method has significant improvement over both the state-of-the-art (SOTA) DL-based methods and the traditional model-based algorithms with less model parameters and less computational load.