Introduction

Computed tomography (CT) is a non-invasive diagnostic imaging tools, known for fast acquisitions, high spatial resolution, and broad availability. Sparse-view CT reconstruction in the dynamic acquisitions is an intriguing yet challenging problem, which involves acquiring only a portion of the full projection data. This low-cost and efficient technique has the benefits, including reducing radiation dose, reducing scan time and improving time-resolving capability in the Cardiac CT [2]. The main challenge in sparse-view reconstruction is angular undersampling that violates Nyquist’s criterion, causing aliasing artifacts [3]. These artifacts manifest as streaks in the CT images and can obscure low-contrast objects of interest, such as lesions. As the number of acquired view projection data decreases, the resulting artifacts become more pronounced and severe. Figure 1 depicts the effect of reducing the sampled data to 1/4, 1/8 and 1/16 of the full data set on the image quality. Such artifacts greatly impact the clinical assessment. Therefore, sophisticated reconstruction algorithms are required to restore the anatomical details and maintain the diagnostic performance. Recent advances in deep learning have shown remarkable progress in various medical imaging tasks, including disease classification, lesion detection, image denoising and artifact removal, etc. The powerful regression capability of the deep neural network enables us to tackle extremely challenging problems. In this study, we introduce a dual-domain convolutional neural network (CNN) to solve the sparse-view CT reconstruction problem.

Related Work

Sparse-view CT reconstruction is an ill-posed inverse problem and there does not exist an analytical solution. In the past, two paradigms have been proposed to recover high-quality images from severely undersampled data. The first paradigm, compressed sensing (CS) [3-5], reformulates an image reconstruction problem to a convex optimization problem with two terms: the data fidelity term, which enforces consistency with the acquired data, and the regularization term, which transforms the image to a new space and promotes the sparsity in that space [3]. This optimization problem can be solved by a range of gradient-descent based methods, such as conjugate gradient and ADMM. In the end, the reconstructed image balances data fidelity and artifact removal [1]. The second paradigm, deep-learning-based methods [1-2,6-10], enables fast and higher-quality image reconstruction compared to CS-based methods. This class of methods can be categorized into three groups: (1) converting artifact-contaminated images to artifact-free images (2) inpainting undersampled projection data to generate a full data set, then applying filter backprojection (FBP), and (3) directly reconstructing images from undersampled projection data without explicit use of the classic reconstruction algorithms (e.g., FBP). In most cases, deep neural networks trained purely on the image domain performed reasonably well, like RED-CNN [11].  However, they did not fully leverage the information from the projection domain. Lack of consistency check with the acquired data may cause false negative lesions and false positive lesion-like structures [1] in the reconstructed images. Similarly, only using the projection data to train the model loses information of spatial correlations in the image domain.

Our Work

In this work, we propose a dual-domain sparse view CT reconstruction pipeline that leverages both sinograms and images. Specifically, we first use a deep neural network to convert the artifact-contaminated CT image to an artifact-reduced image, followed by a Radon transform to convert the image to the sinogram. In the sinogram domain, we develop another model to correct the projection data to obtain a more accurate estimate of the full data set. Finally, we reconstruct the image using FBP. Our contributions are summarized as follows:

• Introduction of a dual-domain CNN to estimate the full projection data.

• Comparison with a purely image-domain model to highlight the importance of incorporating sinogram-domain information.

• Model evaluation on acquisition scenarios different from the training conditions to test for generalizability.