DDPTransformer: Dual-Domain With Parallel Transformer Network for Sparse View CT Image Reconstruction

Abstract

Computed tomography (CT) is increasingly essential for clinical diagnosis nowadays while X-ray ionizing radiation is harmful and may increase the risk of cancers. Researchers have thus proposed to obtain the sparse sinograms by reducing the number of detectors to reduce the radioactive impact on human body. However, the sparse-view sinograms have some problems including blurring and streak artifacts, which may adversely confuse the diagnostic assessment to some extent. In this paper, we propose a Transformer-based module named DDPTransformer Block to capture the long-term dependencies within the images and reconstruct high-quality CT images from sparsely sampled sinograms. More specifically, we replace the traditional convolution operation with the Parallel Transformers, and augment the edge information of patch blocks by different splits. In addition, considering that the images to be processed are two-dimensioned, we propose the Layer-Conv-Layer (LCL) module to replace the multilayer perceptron (MLP) in Transformer for feature extraction. Finally, with DDPTransformer Block as the backbone, we propose a deep learning model with four stages including Interpolation, Sinogram Domain Subnet, Filtered BackProjection (FBP) and Image Domain SubNet. For training the model, we use the dataset from the “ Low Dose CT Image and Projection Data (LDCT-and-Projection data)”. Experimental results show that the proposed model outperforms the state-of-the-art algorithms for various sparsely sampled sinograms and its robustness has been further verified through other datasets. The code and model are publicly available at: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/MikasaLee/DDPTransformer</uri> .