From filtering to denoising: Increasing visual interpretability of cryo-electron tomograms

Cryo-electron tomography has emerged as the premier technique for ultrastructural analysis of natively preserved biological specimens and in situ structure determination. Each tomogram of a cell contains valuable information on the imaged molecular assemblies, leading to potential discoveries, but it also contains enormous amounts of noise. This noise, in combination with the typical low contrast in raw cryo-electron tomograms, hampers the discovery process. To overcome this impairment on the levels of tomogram reconstruction and the reconstructed tomogram, the field has employed a variety of image processing techniques, ranging from binning and low-pass filters removing the typically noisier high-frequency Fourier components to neural network-based denoisers. Here, we provide an overview of the approaches that are used in current research studies and an outlook on a set of newly developed strategies leveraging neural networks to raise the visual interpretability of tomograms and thereby, hopefully, increase the rate of new discoveries.