A strategy to improve image quality of low-dose synchrotron radiation CT imaging for tissue engineering applications
Keywords:Synchrotron-Based Computed Tomography, Radiation dose, Image reconstruction, Deep learning
Hydrogel scaffolds made from biomaterials are used to facilitate cell growth and tissue regeneration and are essential in tissue engineering applications. Hydrogel scaffolds have very low density and synchrotron radiation X-ray computed tomography (SR-CT) shows high contrast for hydrogel scaffolds characterization . However, the radiation dose is of a potential risk using SR-CT for high-resolution imaging in vivo . Reducing the radiation dose has been challenging since low-dose results suffer from noise and artifacts, thus significantly degrading image qualities.
Improving the image quality for low-dose CT imaging has drawn considerable attention over the past decades. Deep learning (DL) methods can help to denoise and remove artifacts; however, most DL-based methods are trained in the supervised mode. Their success critically depends on a large number of paired high-quality data, which typically results in relatively high doses. Such high-quality data are sometimes limited or even impossible to be obtained in biomedical in vivo studies.
We present a low-dose imaging strategy that combines paired high-flux sparse-view CT scan (HF-SV) and low-flux full-view CT scan (LF-FV) based on generative adversarial network (GAN), namely Sparse2Noise. The task is to determine the parameters in network ,
where and are the HF-SV and LF-FV reconstructions, respectively. Sparse2Noise fills the sparse-view artifacts but doesn’t learn the low-dose noise on target.
II. EXPERIMENTS AND RESULTS
We first evaluated Sparse2Noise on the data that were captured from the 3% w/v alginate hydrogel tissue scaffolds in vitro by means of the propagation-based imaging CT (PBI-CT) technique, one of SR-CT techniques. The PBI-CT imaging was performed at the BMIT 05ID-2 beamline, Canadian Light Source, Canada, at a sample-to-detector distance of 1.5 m (30 keV and pixel size of 13 µm). LF-FV images (0.8 Gy) were reconstructed from 1500 projections with 12-cm neutral density filters (NDF). HF-SV images (1.4 Gy) were reconstructed from 75 projections without placing NDF.
Results in Fig. 1 show the noise and ring artifacts are clearly present on LF-FV and HF-SV images, and scaffolds are hard to be identified. By introducing Sparse2Noise, the noise and artifacts can be significantly reduced, thus enhancing the image quality. To produce a similar quality without Sparse2Noise, the dose would be increased to 28.32 Gy.
This study paves the way for in vivo visualization of hydrogel scaffolds using SR-CT with low radiation dose.
Fig. 1 Low-dose SR-CT data of hydrogel scaffolds; (a-c) processed results without and with Sparse2Noise, and (d) the high-dose results as reference.
This work is supported by the Natural Sciences and Engineering Research Council of Canada (Grant numbers: RGPIN 06007-2019 and RGPIN 06396-2019).
 L. Ning et al., "Noninvasive Three-Dimensional In Situ and In Vivo Characterization of Bioprinted Hydrogel Scaffolds Using the X-ray Propagation-Based Imaging Technique," ACS Appl. Mater. Interfaces, vol. 13, no. 22, pp. 25611-25623, 2021.
 K. D. Harrison et al., "Direct Assessment of Rabbit Cortical Bone Basic Multicellular Unit Longitudinal Erosion Rate: A 4D Synchrotron‐Based Approach," J. Bone Miner. Res., vol. 37, no. 11, pp. 2244-2258, 2022.