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

Scanning emission-based microscopies, such as X-ray fluorescence (XRF) and energy-dispersive X-ray spectroscopy, offer nanometer-scale chemical maps, but suffer from long acquisition times and radiation damage. Lower-flux and shorter dwell time scans mitigate this problem, but the resulting signal loss can only partially be compensated by adding detector elements, a strategy limited by the instrument's geometry and cost. We introduce an self-supervised machine learning (ML) pipeline that recovers signal by exploiting the intrinsic redundancy in the data captured by multielement detectors. Inspired by the Noise2Noise approach, we train a deep convolutional neural network directly on the multiple, statistically independent noisy images collected by individual detector elements, without requiring clean training targets. Tested on a resolution target and a cryogenically fixed biological cell, the method markedly improved signal-to-noise ratios over classical filters while preserving spatial resolution and elemental quantification even for small images. To our knowledge, this is the first ML-based denoiser demonstrated for XRF. This approach reduces the dependence of image quality on photon dose and enables rapid, low-dose chemical imaging. It is readily transferable to any modality that records parallel, noise-independent views of the same sample.