A Deep Learning Method for Simultaneous Denoising and Missing Wedge Reconstruction in Cryogenic Electron Tomography

TL;DR

This work tackles denoising and missing wedge reconstruction in cryogenic electron tomography by introducing DeepDeWedge, a self-supervised approach that fits an untrained neural network to tilt-series data. The method follows a three-step pipeline: split the tilt series into two independent-noise sub-tilt-series, train a network with a Noise2Noise–style loss to denoise and fill the wedge, and refine by applying the model and averaging results. Across real datasets (ribosomes, flagella, and ciliary zones) and synthetic SHREC data, DeepDeWedge matches or surpasses state-of-the-art pipelines (CryoCARE + IsoNet, IsoNet, CryoCARE) in denoising, reduces missing wedge artifacts, and yields smoother, higher-contrast tomograms, with 0.143-FSC and CC-based metrics showing favorable performance. The approach minimizes hyperparameter tuning and training data requirements, enabling robust tomogram enhancement, though the authors caution about potential hallucinations and emphasize validation in tricky regions and per-structure trustworthiness.

Abstract

Cryogenic electron tomography is a technique for imaging biological samples in 3D. A microscope collects a series of 2D projections of the sample, and the goal is to reconstruct the 3D density of the sample called the tomogram. Reconstruction is difficult as the 2D projections are noisy and can not be recorded from all directions, resulting in a missing wedge of information. Tomograms conventionally reconstructed with filtered back-projection suffer from noise and strong artifacts due to the missing wedge. Here, we propose a deep-learning approach for simultaneous denoising and missing wedge reconstruction called DeepDeWedge. The algorithm requires no ground truth data and is based on fitting a neural network to the 2D projections using a self-supervised loss. DeepDeWedge is simpler than current state-of-the-art approaches for denoising and missing wedge reconstruction, performs competitively and produces more denoised tomograms with higher overall contrast.

Figures (11)

• Figure 1: Illustration of DeepDeWedge. For simplicity, we show the 2D tilt series images as 1D Fourier slices and all 3D tomograms as 2D objects in the Fourier domain. Recall that tilt series images, tomograms, and sub-tomograms are objects in the image domain. The figure shows the splitting approach where the tilt series is split into even and odd projections.
• Figure 2: Illustration of a single ground-truth structure $\mathbf{v}_{}^*$, and three model input-target pairs corresponding to three different random positions of the original (blue) and additional (red) missing wedge for the setup of our theoretical motivation. For simplicity, we visualize a 2D structure with no additive noise and random in-plane rotations. The inset boxes in each patch show the absolute values of the Fourier transforms of the images and regions that are zeroed out by the missing wedge masks.
• Figure 3: Slices through 3D reconstructions of a tomogram containing purified S. cerevisiae 80S ribosomoes (EMPIAR-10045, Tomogram 5). The red lines in each slice indicate the positions of the remaining two slices. We also show the central x-z-slice through the logarithm of the magnitude of the Fourier transform of each reconstruction.
• Figure 4: Slices through 3D reconstructions of the flagella of C. reinhardtii when using different reconstruction methods. The red lines in each slice indicate the positions of the remaining two slices. We also show the central x-z-slice through the logarithm of the magnitude of the Fourier transform of each reconstruction.
• Figure 5: Slices through 3D reconstructions of tomograms showing the ciliary transit zone of C. reinhardtii. The red lines in each slice indicate the positions of the remaining two slices. We also show the central x-z-slice through the logarithm of the magnitude of the Fourier transform of each reconstruction. The row labels follow the naming convention of EMPIAR-11078.

Theorems & Definitions (4)

• Proposition 1
• Lemma 1
• proof : Proof of Lemma \ref{['lem:Noise2Noise']}
• proof : Proof of Proposition \ref{['prop:selfsupervised_is_supervised']}