Denoising Cryo-Electron Tomograms Using Deep Learning

Abstract

Cryo-Electron Tomography (Cryo-ET) gives structural biologists the tools to inspect biological samples in situ at a near atomic resolution is 3D. The three dimensional volumes called `tomograms' that this process produces suffer from a very low signal-to-noise ratio, and therefore need denoising. This thesis explores 5 neural networks in increasing complexity for their ability to denoise the 3D tomograms. The performance of the network was measured by using multiple metrics including the Structural Similarity (SSIM) score, where 0 is worst and 1 is best. All networks are able to transform noisy tomograms with a mean SSIM score of 0.258 to denoised volumes with a mean SSIM score of 0.937 for the worst network and 0.993 for the best network. The networks show that they generalize very well to unfamiliar particles, moderately well to different noise models, and poorly to multiple particles in a volume. Further research must conclude if the poor performance for the multiple particles is due to the change in scale.