Over time, optical microscopy has improved to allow for increasing resolution. Traditional light microscopes can resolve distances down to 200nm due to the diffraction limit of light. Significant challenges arise when trying to see structures smaller than this.
Single Molecule Localisation Microscopy (SMLM) techniques, includ- ing stochastic optical reconstruction microscopy (STORM) [1], photo- activated localisation microscopy (PALM) [2] and fluorescence photo- activated localisation microscopy (FPALM) [3] have brought nanoscale resolution into biology, circumventing the diffraction limit on micro- scopy by exploiting the photoblinking ability of some fluorophores.
There remain a number of technical challenges involved with SMLM, in both acquisition and analysis of the data. Extracting useful informa- tion from reconstructed images (particularly counting the number of molecules of interest in the sample) is still difficult in samples which are densely-labelled with many overlapping fluorophores. Datasets are also typically very large so simply processing and storing data can be a technical challenge itself.
This thesis postulates that a combination of Gibbs sampling by means of a Mixture Model and a Modified Ising Model (MIM) could be used to improve the accuracy of SMLM reconstructions by characterising each pixel of the image into two populations: signal or noise. Through test- ing on simulated datasets, it was found that this method of classifying pixels can successfully denoise images, and reduce the amount of disk space required to save them by, on average, 95%. Statistical knowledge about the two populations is also determined using a parameterless system.
It was found that using the processed data in combination with machine learning algorithms [4], could provide an improved method for accurately counting fluorophores of interest in sample sections. The accuracy reached values of 88.6%, 89.3%, 90.8% and 98.9%, and loss values of 0.229, 0.213, 0.191 and 0.017 for the non-processed datasets, the Mixture Model output, the MIM output and the product of the Mixture Model and MIM outputs respectively. Using the processed datasets also decreased the number of iterations required for the neural network (NN) to reach a high accuracy/ low loss. These accuracies abstract
and loss values were reached on the training dataset. This method could be used on real SMLM datasets, potentially providing a way to determine the absolute numbers of biological molecules of interest in cells, allowing biologists to extract more quantitative information from SMLM techniques.
The NN was adapted to take in three dimensional data instead of just single images. This data included the previous, current and following images. Unfortunately this only managed to get to an accuracy of 42.4% using the 4 different datasets as before: non-processed data, the Mixture Model output, the MIM output and the product of the Mixture Model and MIM outputs
References
[1] M. Rust, M. Bates and X. Zhuang, ‘Stochastic optical reconstruction microscopy (STORM) provides sub-diffraction-limit image resolu- tion’, Nature methods 3, 793–795 (2006).
[2] E. Betzig, G. H. Patterson, R. Sougrat, O. W. Lindwasser, S. Ole- nych, J. S. Bonifacino, M. W. Davidson, J. Lippincott-Schwartz and H. F. Hess, ‘Imaging intracellular fluorescent proteins at nanometer resolution.’, Science (New York, N.Y.) 313, 1642–5 (2006).
[3] S. T. Hess, T. P. K. Girirajan and M. D. Mason, ‘Ultra-high resolution imaging by fluorescence photoactivation localization microscopy.’, Biophysical journal 91, 4258–72 (2006).
[4] I. Goodfellow, Y. Bengio and A. Courville, Deep learning, http:// www.deeplearningbook.org (MIT Press, 2016).
