Sparse2Noise: Low-dose synchrotron X-ray tomography without high-quality reference data.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Radiology

Background:

• Synchrotron radiation computed tomography (SR-CT) offers high-resolution in vivo imaging but requires high radiation doses.
• Reducing radiation dose by limiting projections or photon flux introduces noise and artifacts.
• Existing deep learning methods for artifact reduction require high-quality reference data, which is often unavailable in low-dose scenarios.

Purpose of the Study:

• To develop a novel low-dose imaging strategy for SR-CT that overcomes the limitations of existing methods.
• To enable high-quality in vivo imaging with reduced radiation exposure.
• To create a method that does not rely on high-quality reference data for training.

Main Methods:

• Introduced Sparse2Noise, a strategy combining sparse-view and full-view CT scan data.
• Utilized a convolutional neural network (CNN) for image reconstruction and denoising.
• Trained the model without requiring high-quality reconstructed data, using small datasets.

Main Results:

• Sparse2Noise effectively reduced noise and ring artifacts, outperforming state-of-the-art denoising methods.
• Achieved high image quality with an acceptable low radiation dose (0.5 Gy) for ex vivo rat hindlimb imaging.
• Demonstrated the capability to train on small datasets without high-quality references.

Conclusions:

• Sparse2Noise represents a significant advancement for in vivo SR-CT imaging.
• The method allows for high-quality imaging at reduced radiation doses.
• Sparse2Noise is applicable to conventional CT and phase-contrast CT denoising.