Shot noise reduction in radiographic and tomographic multi-channel imaging with self-supervised deep learning

Area of Science:

• Medical Imaging
• Image Processing
• Computational Science

Background:

• Shot noise significantly degrades image quality in radiographic and tomographic imaging.
• Reduced signal-to-noise ratio is a critical challenge, particularly under constrained imaging conditions.
• Existing denoising methods often struggle with multi-channel datasets where noise characteristics vary.

Purpose of the Study:

• To develop and validate a self-supervised denoising method for noisy multi-channel imaging datasets.
• To improve image quality by exploiting structural similarities between adjacent image channels.
• To adapt the Noise2Noise approach for applications involving time or energy-resolved imaging.

Main Methods:

• Broadened the application of the Noise2Noise self-supervised denoising algorithm.
• Utilized pairs of samples with identical signals but uncorrelated noise from the data distribution.
• Applied the method to multi-channel datasets where adjacent channels share similar information but have independent noise.
• Demonstrated the method's efficacy through three distinct case studies.

Main Results:

• Successfully improved the quality of noisy multi-channel imaging datasets.
• Validated the method's applicability in spectroscopic X-ray tomography.
• Confirmed performance in energy-dispersive neutron tomography and in vivo X-ray cine-radiography.
• Showcased the effectiveness of exploiting inter-channel structural similarities for noise reduction.

Conclusions:

• The adapted Noise2Noise method effectively denoises multi-channel imaging data by leveraging inter-channel correlations.
• The approach offers a robust solution for enhancing image quality in various time- and energy-resolved imaging modalities.
• This work advances self-supervised learning techniques for critical applications in medical and scientific imaging.