Enhancing synchrotron radiation micro-CT images using deep learning: an application of Noise2Inverse on bone imaging

Area of Science:

• Biomedical Engineering
• Materials Science
• Radiology

Background:

• In situ synchrotron radiation micro-computed tomography (SRµCT) is crucial for studying bone mechanical properties and microstructure.
• Low-dose CT imaging minimizes radiation damage but introduces noise, hindering traditional analysis.
• Self-supervised deep learning offers a potential solution for noise reduction in low-dose imaging.

Purpose of the Study:

• To evaluate the effectiveness of the Noise2Inverse deep learning method for noise reduction in low-dose SRµCT bone imaging.
• To assess the impact of Noise2Inverse on bone microstructure quantification at various simulated radiation doses.
• To identify limitations and potential applications of Noise2Inverse in in situ bone imaging experiments.

Main Methods:

• Simulated low-dose SRµCT datasets were generated by downsampling projection data from full dose to one-sixth frequency.
• The Noise2Inverse deep learning model was trained and applied to denoise these simulated datasets.
• Image quality was assessed visually, and bone microstructural features (lacunae volume, aspect ratio, mineralization) were quantified.

Main Results:

• Noise2Inverse visually recovered high image quality across all simulated dose levels.
• Minor shifts in microstructural parameters were observed at higher doses (full, 1/2, 1/3), while significant distortions occurred at lower doses (1/4, 1/6).
• Training with a larger dataset revealed significant microstructural changes even at a 1/3 dose, suggesting the need for validation scans.

Conclusions:

• Noise2Inverse shows promise for noise reduction in low-dose SRµCT bone imaging, improving visual quality.
• Quantification of bone microstructure using this method requires careful consideration of radiation dose and potential parameter distortions.
• Noise from the experimental setup is a key factor affecting the viability of Noise2Inverse, highlighting the need for optimized imaging parameters and dose calculations.