Multi-stage Deep Learning Artifact Reduction for Pallel-beam Computed Tomography

TL;DR

This work tackles artifact-ridden synchrotron CT data by introducing a three-stage deep learning pipeline that performs artifact reduction in the projection, sinogram, and reconstruction domains. Each stage employs a separate 2D CNN with bypass connections that incorporate both raw data and prior stage outputs to mitigate error propagation, while training proceeds independently for computational efficiency. Across simulated and real-world datasets (Foam, LoDoInd, Tomobank), the method achieves superior PSNR and SSIM compared with classical plus post-processing and end-to-end post-processing approaches, and does so with favorable computation times. The approach integrates smoothly with existing CT pipelines and shows promise for extension to other imaging modalities and self-supervised training regimes.

Abstract

Computed Tomography (CT) using synchrotron radiation is a powerful technique that, compared to lab-CT techniques, boosts high spatial and temporal resolution while also providing access to a range of contrast-formation mechanisms. The acquired projection data is typically processed by a computational pipeline composed of multiple stages. Artifacts introduced during data acquisition can propagate through the pipeline, and degrade image quality in the reconstructed images. Recently, deep learning has shown significant promise in enhancing image quality for images representing scientific data. This success has driven increasing adoption of deep learning techniques in CT imaging. Various approaches have been proposed to incorporate deep learning into computational pipelines, but each has limitations in addressing artifacts effectively and efficiently in synchrotron CT, either in properly addressing the specific artifacts, or in computational efficiency. Recognizing these challenges, we introduce a novel method that incorporates separate deep learning models at each stage of the tomography pipeline-projection, sinogram, and reconstruction-to address specific artifacts locally in a data-driven way. Our approach includes bypass connections that feed both the outputs from previous stages and raw data to subsequent stages, minimizing the risk of error propagation. Extensive evaluations on both simulated and real-world datasets illustrate that our approach effectively reduces artifacts and outperforms comparison methods.

Figures (10)

• Figure 1: Schematic representation of a typical synchrotron CT pipeline, consisting of three stages: projection, sinogram, and reconstruction stage. Each stage may consist of multiple processing steps.
• Figure 2: Comparison of deep learning-based post-processing performance on reconstructed images from the simulated foam phantom dataset pelt2022foam affected by local artifacts (noise) and non-local artifacts. The red and green insets show enlarged views of the affected areas. Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) wang2004image values are provided in the top-right and lower-right corners, respectively. Deep learning-based post-processing is less effective in reducing global artifacts than noise. While this figure demonstrates differences in artifact reduction based on simulation, similar challenging scenarios can occur in real-world dynamic experiments buhrer2019high or fast scan setups raufaste2015threemokso2017gigafrost at synchrotron facilities.
• Figure 3: Representations of noise, ring, and zinger artifact in projection, sinogram, and reconstruction stages. Red patterns are schematic illustrations of distortions. Noise is a local artifact in images of all stages. Distorted pixel values in the same positions of projection images become a line in the sinogram, resulting in a ring artifact in the reconstructed image. Extremely high pixel values (randomly distributed in different positions) remain as high-value spots in the sinogram and cause crossing lines in the reconstructed image as zinger artifacts.
• Figure 4: Schematic of our proposed multi-stage artifact reduction method, illustrating distortions and their correction at each stage. Our method includes bypass connections that incorporate both raw and previously processed data at each stage to reduce the risk of error propagation throughout the pipeline.
• Figure 5: Example image with various levels of artifacts on a simulated foam phantom dataset. The artifacts, including noise, ring artifacts, and zinger artifacts, were generated by varying the parameters $I_0$, $P_{\text{ring}}$, and $P_{\text{zinger}}$, respectively. The PSNR and SSIM metrics with respect to the ground truth image are provided for each reconstructed image, displayed in the bottom left.