# **Summary:**
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3314964/

**Theory:**

The current iterative method of reconstructing 3d fourier spaces
from 2d images (single particle reconstruction), is prone to
overfitting due to the high signal-to-noise ratio present in
Cryo-EM images. Current methods to counter this, such as applying
smoothness filtering, are largely based on heuristics and thus
involve arbitrary decisions.

To avoid this, one can use a statistical framework. Maximum
Likelihood (ML) equations have shown to be effective, however
they rely on large datasets being present, and do not take
into account prior knowledge. As opposed to this, Bayesian
Maximum A Posteriori (MAP) estimation utilizes previous data
of smoothness, and does not rely on a large existing dataset.

This paper utilises this MAP estimation technique with a Gaussian
prior of smoothness in order to correctly estimate the level of
smoothness at every resolution. A Gaussian distribution of noise
is a valid assumption, as conventional techniques make the same
assumption.

**Summary of Results:**

MAP was tested in against the existing conventional
classification library, XMIPP, on 3 different datasets. MAP
resulted in lower overfitting and greater information accuracy
across all resoutions, especially when the existing dataset was
insufficient. Some heuristics were employed when determining the
characteristsics of the Gaussian smoothness prior. Although heuristics
were employed, it was to a much lesser degree than the conventional
method.

In the future, the Bayesian framework can use data
other than smoothness for the prior information input. Overall,
the amount of arbitrariness was reduced and objectivity was increased.

# **Questions:**

Is using a Bayesian framework more/less computationally expensive?

Could a Bayesian approach potentially result in oversmoothening
on a very comprehensive dataset?