Denoising Synchrotron Data at Scale: Balancing Effectiveness and Ease of Use

• Anass BELLACHEHAB

Synchrotron experiments generate large volumes of data with diverse sources of noise, making image denoising a crucial task for obtaining high-quality images. Traditional denoising techniques have been developed and applied in this context, but modern machine learning-based methods have shown great promise in achieving higher denoising performance. In this talk, we review the challenges of denoising in the light of the diversity and size of synchrotron data at scale and evaluate the effectiveness of traditional denoising techniques and modern machine learning-based methods in handling these challenges. We also emphasize the importance of user-friendliness and ease of appropriation of the denoising tools, particularly for machine learning-based methods. We break down the talk into three parts: image denoising in the context for spectral inverse space, image denoising in the real space like tomography, and attempts at inverse problems (missing angle, compressed images). For each of these areas, we provide an overview of common sources of noise, review traditional denoising techniques and modern machine learning-based methods, evaluate their effectiveness in handling diverse synchrotron data at Soleil, and discuss their user-friendliness and ease of adoption. The talk aims to provide insights into the importance of balancing denoising effectiveness and ease of use, particularly for modern machine learning-based denoising techniques, in synchrotron data. We conclude with suggestions for further research and development of user-friendly and accessible denoising techniques, with a focus on modern machine learning-based methods, for synchrotron data.

Machine learning denoising of high-resolution nanotomography data

• Silja Flenner (Hereon)

Hard X-ray nanotomography is a commonly used tool in many research areas such as material science, biology and medicine. Noise is one of the major limitations for full-field nanotomography, in particular when a high time resolution is required. To obtain a high image quality even at high temporal resolution, filtering techniques have to be considered. However, due to the high and often non-uniform noise level (e.g. caused by inhomogeneous illumination) in transmission X-ray microscopy (TXM), the noise reduction without loss of spatial resolution is challenging. In the recent years, there has been a lot of progress in image filtering techniques. In particular machine learning (ML) approaches specialized on tomography data have a high potential to improve the image quality. Several approaches have been made using ML for denoising tomographic data. Yang et al. and Pelt et al. showed that ML can be used to perform fast tomography scans and reduce the noise afterwards. These approaches however require a certain scanning protocol and cannot be used on data acquired without this protocol. In this talk, an approach based on machine learning applied to standard synchrotron nanotomography data will be presented. The projections of a TXM scan are split into two independent stacks which are reconstructed separately. The resulting two reconstructions are from the same measurement and one is used as the input and the other one as the reference for the ML training. This concept (Noise2Inverse) has very recently been proven mathematically by Hendriksen et al. The self-supervised denoising ML technique can be used in a very efficient way to eliminate noise from nanotomography data. The technique presented here is applied to high-resolution nanotomographic data and compared to conventional filters such as a median filter and a non-local means filter optimized for tomographic data sets. The ML approach proves to be a very powerful tool that outperforms conventional filters by eliminating noise without blurring relevant structural features, enabling efficient quantitative analysis in different scientific fields.

Discussion

• Peter Steinbach (HZDR)