User:Acyrusty/Transmission electron cryomicroscopy

Image Processing in Cryo-TEM
Even though in the majority of approaches in electron microscopy one tries to get the best resolution image of the material, it is not always the case in cryo-TEM. Besides all the benefits of high resolution images, the signal to noise ratio remains the main hurdle that prevents assigning orientation to each particle. For example, in macromolecule complexes, there are several different structures that are being projected from 3D to 2D during imaging and if they are not distinguished the result of image processing will be blurry. That is why the probabilistic approaches become more powerful in this type of investigation. There are two popular approaches that are widely used nowadays in cryo-EM image processing, the maximum likelihood approach that was discovered in 1998 and relatively recently adapted Bayesian approach.

The maximum likelihood estimation approach comes from statistics and it deals with probability distributions of the observed particles, summing up all the possible orientations, while in the typical least square estimation, particles get exact orientations per image. This way, the particles in the sample get "fuzzy" orientations after calculations, weighted by corresponding probabilities. The whole process is iterative and aims at optimizing the initial model as best as possible. If the data does not have a lot of noise and the particles do not have any preferential direction, the optimization converges fast and the following image comes close to the real structure. However, in maximum likelihood approach the result depends on the initial guess and sometimes can stop at local minimum.

The Bayesian approach that is now being used in cryo-TEM is empirical by nature. This means that the given dataset is used as a basis for inferring the distribution of particles where in the usual Bayesian method there is a fixed prior probability that is changed after the data is observed. The main difference from the maximum likelihood estimation lies in special reconstruction term that helps smoothing the resulting maps while also decreasing the noise during reconstruction. The smoothing of the maps occurs through assuming the first, prior probability to be a Gaussian distribution and analyzing the data in the Fourier space. Since the connection between the prior knowledge and the dataset is established, there is less chance for human factor errors which potentially increases the objectivity of image reconstruction.

With emerging new methods of cryo-TEM imaging and image reconstruction the new software solutions appear that help to automate the process. After the empirical Bayesian approach have been implemented in the open source computer program RELION (REgularized LIkelihood OptimizatioN) for 3D reconstruction, the program became widespread in the cryo-TEM field. It offers a range of corrections that improve the resolution of reconstructed images, allows implementing versatile scripts using python language and executes the usual tasks of 2D/3D model classifications or creating de novo models.