Suppressing electron exposure artifacts: An electron scanning paradigm with Bayesian machine learning

Electron microscopy of biological, polymeric, and other beam-sensitive structures is often hampered by deleterious electron beam interactions. In fact, imaging of such beam-sensitive materials is limited by the allowable radiation dosage rather that capabilities of the microscope itself, which has been compounded by the availability of high brightness electron sources. Reducing dwell times to overcome dose-related artifacts, such as radiolysis and electrostatic charging, is challenging due to the inherently low contrast in imaging of many such materials. These challenges are particularly exacerbated during dynamic time-resolved, fluidic cell imaging, or three-dimensional tomographic reconstruction - all of which undergo additional dosage. Thus, there is a pressing need for the development of techniques to produce high-quality images at ever lower electron doses. In this contribution, we demonstrate direct dose reduction and suppression of beam-induced artifacts through under-sampling pixels, by as much as 80% reduction in dosage, using a commercial scanning electron microscope with an electrostatic beam blanker and a dictionary learning in-painting algorithm. This allows for multiple sparse recoverable images to be acquired at the cost of one fully sampled image. We believe this approach may open new ways to conduct imaging, which otherwise require compromising beam current and/or exposure conditions.