F2FD: Fourier Perturbations for Denoising Cryo-Electron Tomograms and Comparison to Established Approaches

Cryo-electron tomography (Cryo-ET) is an imaging technique capable of visualizing vitrified biological samples at sub-nanometer resolution in 3D. However, beam-induced damage limits the applied electron dose and leads to a low signal-to-noise ratio. A popular method for denoising cryo-electron tomograms is Cryo-CARE, which performs noise-to-noise training, which relies on splitting the 2D tilt series into two separate halves. In practice, often the original tilt series is not available, but only the reconstructed volume, to which Cryo-CARE cannot be applied. More general denoising methods such as Noise2Void (N2V) or Self2Self with Dropout (S2Sd) do not require noisy image pairs and work with single noisy inputs. However, these methods implicitly assume noise to be pixel-independent, which is not the case for tomographic reconstructions. We propose F2Fd, a deep learning denoising algorithm that can be applied directly to reconstructed tomograms. F2Fd creates paired noisy patches by perturbing high frequencies in Fourier space and performs noise-to-noise training with them. We benchmark F2Fd with five other state-of-the-art denoising methods (including N2V, S2Sd and Cryo-CARE) on both synthetic and real tomograms. We show that the perturbation in Fourier space is better suited for Cryo-ET noise than noise from real space used by N2V and S2Sd. Moreover, we illustrate that Cryo-ET denoising not only leads to cleaner images, but also facilitates membrane segmentation as an important downstream task.