Self-Supervised Deep-Learning Denoising for X-ray Fluorescence Microscopy with Multi-Element Detectors.

Area of Science:

• Microscopy
• Machine Learning
• Spectroscopy

Background:

• Scanning emission-based microscopies like X-ray fluorescence (XRF) provide nanoscale chemical maps.
• However, these techniques are limited by long acquisition times and potential radiation damage.
• Reducing scan times and flux leads to signal loss, which is difficult to fully recover.

Purpose of the Study:

• To develop a novel machine learning (ML) pipeline for signal recovery in XRF.
• To improve signal-to-noise ratios and enable rapid, low-dose chemical imaging.
• To reduce the dependence of image quality on photon dose.

Main Methods:

• A self-supervised deep convolutional neural network (CNN) was trained using the Noise2Noise approach.
• The CNN exploited intrinsic data redundancy from multielement detectors.
• The model was trained directly on statistically independent, noisy images without requiring clean training targets.

Main Results:

• The ML pipeline significantly improved signal-to-noise ratios compared to classical filters.
• Spatial resolution and elemental quantification were preserved, even for small images.
• Demonstrated effectiveness on a resolution target and a biological cell sample.

Conclusions:

• This work presents the first ML-based denoiser for XRF, significantly enhancing data quality.
• The approach enables faster and lower-dose chemical imaging, overcoming key limitations of current methods.
• The technique is transferable to other modalities capturing parallel, noise-independent views.