Machine-learning approach for quantified resolvability enhancement of low-dose STEM data

High-resolution electron microscopy is achievable only when a high electron dose is employed, a practice that may cause damage to the specimen and, in general, affects the observation. This drawback sets some limitations on the range of applications of high-resolution electron microscopy. Our work proposes a strategy, based on machine learning, which enables a significant improvement in the quality of Scanning Transmission Electron Microscope images generated at low electron dose, strongly affected by Poisson noise. In particular, we develop an autoencoder, trained on a large database of images, which is thoroughly tested on both synthetic and actual microscopy data. The algorithm is demonstrated to drastically reduce the noise level and approach ground-truth precision over a broad range of electron beam intensities. Importantly, it does not require human data pre-processing or the explicit knowledge of the dose level employed and can run at a speed compatible with live data acquisition. Furthermore, a quantitative unbiased benchmarking protocol is proposed to compare different denoising workflows.