Measuring cryo-TEM sample thickness using reflected light microscopy and machine learning.

In cryo-transmission electron microscopy (cryo-TEM), sample thickness is one of the most important parameters that governs image quality. When combining cryo-TEM with other imaging methods, such as light microscopy, measuring and controlling the sample thickness to ensure suitability of samples becomes even more critical due to the low throughput of such correlated imaging experiments. Here, we present a method to assess the sample thickness using reflected light microscopy and machine learning that can be used prior to TEM imaging of a sample. The method makes use of the thin-film interference effect that is observed when imaging narrow-band LED light sources reflected by thin samples. By training a neural network to translate such reflection images into maps of the underlying sample thickness, we are able to accurately predict the thickness of cryo-TEM samples using a light microscope. We exemplify our approach using mammalian cells grown on TEM grids, and demonstrate that the thickness predictions are highly similar to the measured sample thickness. The open-source software described herein, including the neural network and algorithms to generate training datasets, is freely available at github.com/bionanopatterning/thicknessprediction. With the recent development of in situ cellular structural biology using cryo-TEM, there is a need for fast and accurate assessment of sample thickness prior to high-resolution imaging. We anticipate that our method will improve the throughput of this assessment by providing an alternative method to screening using cryo-TEM. Furthermore, we demonstrate that our method can be incorporated into correlative imaging workflows to locate intracellular proteins at sites ideal for high-resolution cryo-TEM imaging.