End-to-End Deep Learning for Interior Tomography with Low-Dose X-ray CT

TL;DR

This work tackles radiation-dose reduction for CT by combining interior ROI tomography with low-dose strategies, a setting that induces coupled cupping artifacts and image noise challenging for conventional reconstruction. It proposes an end-to-end dual-domain CNN architecture: a projection-domain CNN first decouples inside-ROI projection noise and outside-ROI extrapolation, followed by an image-domain CNN, connected via a differentiable FBP layer and solved in an unrolled framework. Grounded in deep convolutional framelets theory, the method leverages low-rank Hankel structures to justify projection-domain processing and yields improved NMSE and SSIM over image-domain DL and MBIR baselines across varied ROI sizes and dose levels. The results indicate that projection-domain decoupling enhances artifact suppression while preserving texture, enabling more accurate and efficient low-dose interior CT reconstructions with practical clinical impact.

Abstract

Objective: There exist several X-ray computed tomography (CT) scanning strategies to reduce a radiation dose, such as (1) sparse-view CT, (2) low-dose CT, and (3) region-of-interest (ROI) CT (called interior tomography). To further reduce the dose, the sparse-view and/or low-dose CT settings can be applied together with interior tomography. Interior tomography has various advantages in terms of reducing the number of detectors and decreasing the X-ray radiation dose. However, a large patient or small field-of-view (FOV) detector can cause truncated projections, and then the reconstructed images suffer from severe cupping artifacts. In addition, although the low-dose CT can reduce the radiation exposure dose, analytic reconstruction algorithms produce image noise. Recently, many researchers have utilized image-domain deep learning (DL) approaches to remove each artifact and demonstrated impressive performances, and the theory of deep convolutional framelets supports the reason for the performance improvement. Approach: In this paper, we found that the image-domain convolutional neural network (CNN) is difficult to solve coupled artifacts, based on deep convolutional framelets. Significance: To address the coupled problem, we decouple it into two sub-problems: (i) image domain noise reduction inside truncated projection to solve low-dose CT problem and (ii) extrapolation of projection outside truncated projection to solve the ROI CT problem. The decoupled sub-problems are solved directly with a novel proposed end-to-end learning using dual-domain CNNs. Main results: We demonstrate that the proposed method outperforms the conventional image-domain deep learning methods, and a projection-domain CNN shows better performance than the image-domain CNNs which are commonly used by many researchers.

Figures (10)

• Figure 1: Low-dose ROI CT compositions of (a) measurement $p_{\mathscr{T}} = \mathscr{T} \odot y + h_{\mathscr{T}}$ in projection domain and (b) FBP image $q_{\mathscr{I}} = f_{\mathscr{I}} + c_{\mathscr{I}} + n_{\mathscr{I}}$ in image domain.
• Figure 2: Various neural network architectures. (a) image-domain CNN, (b) projection-domain CNN, (c) W-Net, and (d) proposed network (called Dual-Net). (e) describes function modules used in (a-d).
• Figure 3: A CT coordinate system.
• Figure 4: (a) Image noise property of low-dose CT and (b) Cupping artifact property of ROI CT. $\mathscr{F}$ denotes 2D Fourier transform.
• Figure 5: (a) A backbone based on the standard U-Net structure, (b) image-domain CNN $\mathscr{Q}_{img}$ consisting a backbone and a single bridge module to estimate an image, (c) projection-domain CNN $\mathscr{Q}_{prj}$ consisting the backbone and two bridge modules to estimate a projection noise inside the measured region $\mathscr{T}$ and an extrapolation map outside the measured region $(1 - \mathscr{T})$, respectively, and (d) a bridge module. (e) shows definition of layers.