[Paper Review] On the Benefit of Dual-domain Denoising in a Self-supervised Low-dose CT Setting

Fabian Wagner, Mareike Thies+10 more|arXiv (Cornell University)|Nov 2, 2022

Medical Imaging Techniques and Applications4 citations

TL;DR

This paper proposes an end-to-end, self-supervised, dual-domain denoising pipeline for low-dose CT that jointly optimizes denoising in both projection and image domains without requiring high-dose ground truth. It achieves up to 94.1% PSNR improvement on abdomen CT and 2.9% PSNR gain on XRM data over low-dose baselines, demonstrating that early noise suppression in projection space significantly enhances final image quality.

ABSTRACT

Computed tomography (CT) is routinely used for three-dimensional non-invasive imaging. Numerous data-driven image denoising algorithms were proposed to restore image quality in low-dose acquisitions. However, considerably less research investigates methods already intervening in the raw detector data due to limited access to suitable projection data or correct reconstruction algorithms. In this work, we present an end-to-end trainable CT reconstruction pipeline that contains denoising operators in both the projection and the image domain and that are optimized simultaneously without requiring ground-truth high-dose CT data. Our experiments demonstrate that including an additional projection denoising operator improved the overall denoising performance by 82.4-94.1%/12.5-41.7% (PSNR/SSIM) on abdomen CT and 1.5-2.9%/0.4-0.5% (PSNR/SSIM) on XRM data relative to the low-dose baseline. We make our entire helical CT reconstruction framework publicly available that contains a raw projection rebinning step to render helical projection data suitable for differentiable fan-beam reconstruction operators and end-to-end learning.

Motivation & Objective

• To address the limitation of existing self-supervised CT denoising methods that operate only in the image domain, missing early noise suppression in projection data.
• To develop an end-to-end trainable pipeline that jointly optimizes denoising in both projection and image domains without requiring paired high-dose data.
• To demonstrate that dual-domain denoising leads to superior image quality compared to single-domain post-processing denoising.
• To release a fully differentiable, open-source framework for helical CT data rebinning and reconstruction to enable future research on dual-domain and self-supervised CT pipelines.

Proposed method

• The method uses a differentiable reconstruction operator R and two trainable denoising modules: D_proj^w in the projection domain and D_img^w in the image domain.
• The pipeline is trained end-to-end via backpropagation through the entire chain: D_proj^w → R → D_img^w → loss, enabling joint optimization of both denoisers.
• A self-supervised loss is computed between two independently reconstructed views from split projection stacks, enabling training without high-dose ground truth.
• The framework includes a raw projection rebinning step to convert helical CT data into fan-beam compatible format for differentiable reconstruction.
• The method supports arbitrary differentiable denoising architectures (e.g., U-Nets, BFs) in both domains, enabling flexible integration.
• The entire pipeline is trained using a contrastive loss based on feature similarity between two denoised reconstructions from independent projection splits.

Fig. 1 : Overview of the proposed trainable dual-domain and self-supervised CT denoising pipeline.

Experimental results

Research questions

• RQ1 Does incorporating denoising in the projection domain improve low-dose CT reconstruction performance compared to image-domain-only denoising?
• RQ2 Can a self-supervised, end-to-end trainable pipeline jointly optimize denoising in both projection and image domains without high-dose ground truth?
• RQ3 How does dual-domain denoising compare quantitatively and qualitatively to single-domain post-processing denoising on medical and pre-clinical CT data?
• RQ4 To what extent does early noise suppression in the projection domain simplify the subsequent image denoising task?
• RQ5 Can a differentiable, open-source framework for helical CT rebinning and reconstruction enable broader adoption of dual-domain, self-supervised CT methods?

Key findings

• Dual-domain denoising improved PSNR by 82.4–94.1% and SSIM by 12.5–41.7% on abdomen CT data relative to the low-dose baseline.
• On X-ray microtomography (XRM) data, the method achieved a 1.5–2.9% PSNR gain and 0.4–0.5% SSIM improvement over the low-dose baseline.
• The dual-domain self-supervised pipeline outperformed all single-domain post-processing baselines, confirming that early denoising in the projection domain is beneficial.
• The proposed framework achieved performance close to supervised baselines, with the dual U-Net self-supervised model reaching 32.5 dB PSNR on XRM data.
• Visual inspection confirmed that dual-domain models reduced noise more effectively and preserved anatomical details better than image-only denoising, especially in regions with low contrast like liver lesions.
• The open-source helical CT rebinning and differentiable reconstruction framework enables end-to-end learning on raw projection data, facilitating future research in dual-domain and self-supervised CT.

Fig. 2 : Illustration of the proposed end-to-end self-supervised denoising pipeline. Projections are split in two stacks and reconstructed separately. The loss calculated between denoised prediction and reconstructed target stack is backpropagated through $\mathbf{D}_{\text{img}}^{w}$ and $\mathbf{R