Deep denoising for multi-dimensional synchrotron X-ray tomography without high-quality reference data.

Area of Science:

• Materials Science
• Image Processing
• Computational Imaging

Background:

• Synchrotron X-ray tomography provides high-resolution material structure insights but is limited by noise from dose/time constraints.
• Convolutional Neural Networks (CNNs) are effective for image denoising but typically require paired noisy and high-quality training data.
• Existing CNN denoising methods often necessitate extensive, specific datasets, posing a practical challenge for tomography.

Purpose of the Study:

• To adapt and expand the Noise2Inverse method for denoising synchrotron X-ray tomography data.
• To enable effective image quality enhancement without requiring separate training datasets.
• To broaden the applicability of Noise2Inverse to static, dynamic, and diffraction-based tomography.

Main Methods:

• Extension of the Noise2Inverse method to process data across spatial, temporal, and spectral domains.
• Application of the adapted Noise2Inverse method to real-world synchrotron X-ray tomography datasets.
• Evaluation of denoising performance and impact on acquisition time and image quality.

Main Results:

• The enhanced Noise2Inverse method successfully denoises reconstructed images from synchrotron X-ray tomography.
• Significant reduction in data acquisition time was achieved while preserving high image quality.
• The method demonstrated effectiveness across static, dynamic, and X-ray diffraction tomography applications.

Conclusions:

• The expanded Noise2Inverse method offers a powerful, data-efficient solution for denoising in various synchrotron tomography techniques.
• This approach overcomes the limitations of traditional training data requirements for CNN-based denoising.
• The findings enable faster, higher-quality material characterization using synchrotron X-ray tomography.