Why not electron microscopy if machine learning is meant to make everything easier for us these days? Scanning transmission electron microscopy has been improved using machine learning by a team from Ireland, and they published their findings. The outcome is significant because it caters to a very specific use case—low dose STEM.
The issue is that high electron doses are frequently required to obtain high resolutions. However, exposing a sensitive, frequently biological, topic to a barrage of high-energy electrons could alter the results and harm the sample. Due to Poisson noise, however, employing decreased electron doses produces a poor image. Even at modest dosages, the new method picks up on how to account for the noise and produce a higher-quality image.
Because it moves quickly enough to function in real time, the processing doesn’t need human interaction. The tiny features in the scans shown in the paper are difficult to read, but it is clear that the usual Gaussian filter does not perform as well. Following filtering, the first dots seem “fat.” The new method draws attention to the tiny dots while reducing background noise. This is one of those tasks that can be completed by a human with relative ease, yet typical computer methods don’t necessarily yield the best outcomes.
You have to wonder what other aspects of signal processing machine learning could enhance. Naturally, you want to make sure that you aren’t creating data that isn’t there. Teaching a CAT scan computer that everyone requires an expensive surgery wouldn’t be wise. That is unimaginable, right?
It is challenging to shoot an electron through an electronic component, hence in electronic work we typically utilize SEM that detects secondary electron emissions. But STEM is a fantastic technology that even allows for the display of atoms’ shadows. It is still a rather huge trek to have an electron microscope in your home lab, but we keep hoping that someone will develop a homebrew design that would be simple to copy.
