Cross-domain Denoising for Low-dose Multi-frame Spiral Computed Tomography

TL;DR

A two-stage method for the commercially available multi-slice spiral CT scanners that better exploits the complete reconstruction pipeline for LDCT denoising across different domains that makes good use of the high redundancy of multi-slice projections and the volumetric reconstructions while leveraging the over-smoothing issue in conventional cascaded frameworks caused by aggressive denoising.

Abstract

Computed tomography (CT) has been used worldwide as a non-invasive test to assist in diagnosis. However, the ionizing nature of X-ray exposure raises concerns about potential health risks such as cancer. The desire for lower radiation doses has driven researchers to improve reconstruction quality. Although previous studies on low-dose computed tomography (LDCT) denoising have demonstrated the effectiveness of learning-based methods, most were developed on the simulated data. However, the real-world scenario differs significantly from the simulation domain, especially when using the multi-slice spiral scanner geometry. This paper proposes a two-stage method for the commercially available multi-slice spiral CT scanners that better exploits the complete reconstruction pipeline for LDCT denoising across different domains. Our approach makes good use of the high redundancy of multi-slice projections and the volumetric reconstructions while leveraging the over-smoothing problem in conventional cascaded frameworks caused by aggressive denoising. The dedicated design also provides a more explicit interpretation of the data flow. Extensive experiments on various datasets showed that the proposed method could remove up to 70\% of noise without compromised spatial resolution, and subjective evaluations by two experienced radiologists further supported its superior performance against state-of-the-art methods in clinical practice.

Figures (15)

• Figure 1: Overview of the proposed multi-stage hierarchical framework. The curly brackets indicate the concatenation operation. Note that only a single stream is shown in the projection-domain denoising, and all the noise components are amplified for better visibility.
• Figure 2: A sample clip from the raw projections obtained by a multi-slice spiral CT scanner.
• Figure 3: The structure of MPD-Net: (a) MPD-Net with $F$=2, where ResUNets with different colors have individual parameter sets; (b) the structure of ResUNet. Details of E1-E2 and D1-D2 are given in Fig. \ref{['fig-mirnet']}.
• Figure 4: Two examples of inter-slice (study: Siemens-L291) and intra-slice (study: Siemens-L291, slice: 9,500) dose levels. It can be seen that the dose level not only varies dramatically within the scan but also has a non-uniform distribution across detector columns.
• Figure 5: Structure of MIR-Net. The number in each layer represents the output channel size, and $\bigoplus$ represents the adaptive mix-up operation. E1-E2 and D1-D2 represent encoding and decoding blocks, respectively. All layers use ReLU as the activation function.