Denoising of electron tomograms

Abstract

The crucial problem inherent to electron tomography is radiation damage or, related to this, the choice of the correct electron dose: an excessive dose destroys the specimen, especially biological ones, while an insufficient dose results in images that are noisy and lack information. Sophisticated and highly automated techniques have been developed both for data acquisition with the aim of keeping the electron dose as low as possible, and for image processing, in order to extract reliable information from the recorded data. However, the tolerable dose is very small, especially for unstained, frozen-hydrated specimens. As a rule of thumb, 5000e/nm2 are tolerable for such specimens. According to the dose fractionation theorem (Hegerl and Hoppe, 1978), the total tolerable dose has to be divided by the number of projection views in order to find the dose allowed for each image of a tilt series. In addition, the low scattering power of biological material results in low-contrast images. For instance, assuming a tilt series of 50 images, a pixel size of 1nm2, phase contrast imaging with a contrast of 10%, and considering only the shot noise of the electrons, the signal-to-noise ratio (SNR defined as energy of signal over energy of noice) in the projection images is in the order of 1. An increase in the number of projection images, a decrease of the pixel size and additional noise arising from the image recording system push the SNR below 1.