Machine learning is supposed to help us do just about everything these days, but why electron microscopy? published results that enhance electron microscopy). This result is important because it targets a very specific use case: low-dose STEM.
The problem is that obtaining high resolution usually requires the use of high electron doses. However, bombarding delicate, often biological objects with high-energy electrons can change what you see and damage your sample. However, reducing the electron dose used reduces image quality due to Poisson noise. This new technique learns how to correct for noise and produce higher quality images at lower doses.
The processing does not require human intervention and is fast enough to work in real time. Small features of the scans presented in the paper are difficult to interpret, but it turns out that the standard Gaussian filter does not work as well. The original dots appear “fat” after filtering. New technology emphasizes small dots and reduces noise between them. It’s one of those things that humans can easily do, but traditional computer technology doesn’t always give great results.
We need to think about what else machine learning can improve on signal processing. Of course, you have to make sure you’re not creating data that doesn’t exist. He can’t tell the CAT scan computer that everyone needs an expensive surgical procedure. It never happens, right?
Electronic engineering typically uses SEMs to detect secondary electron emissions because it is difficult to emit electrons through electronic components. However, STEM is an excellent technology that can display atomic shadows uniformly. I keep hoping that someone will come up with a home-made design that can be easily replicated, but so far it’s been a pretty long way to get an electron microscope for my home lab.
