A Monte Carlo framework for missing wedge restoration and noise removal in cryo-electron tomography

Emmanuel Moebel, Charles KervrannJournal of Structural Biology X 2019 · DOI

TL;DR

This paper introduces a novel Monte Carlo Markov Chain (MCMC) framework called MWR (Missing Wedge Restoration) to tackle the pervasive "missing wedge" artifact and noise in cryo-electron tomography (cryo-ET). By employing a statistical approach to estimate the uncorrupted image, MWR effectively fills in missing spectral information, leading to significantly improved 3D reconstructions for biological imaging.

What Problem Does This Paper Address?

Cryo-electron tomography (cryo-ET) is a powerful technique for visualizing cellular structures at high resolution. However, it suffers from two major limitations: (1) high noise levels due to the need to minimize electron dose to preserve the sample, and (2) the "missing wedge" artifact. The latter arises because it's impossible to tilt the sample a full 90 degrees, leading to a gap in the collected projection data. This gap manifests as a "missing wedge" in the Fourier domain, causing artifacts like ray and side streaks, and object elongation in the reconstructed 3D volumes. These artifacts severely hinder the interpretation and analysis of biological structures. Existing methods often struggle to effectively compensate for both noise and the missing wedge simultaneously, especially in complex biological samples.

Methods

The core of this work is a statistical method based on a Bayesian framework to estimate the Minimum Mean Square Error (MMSE) of the uncorrupted image. To handle the high-dimensional integral required for MMSE estimation, the authors propose a Markov Chain Monte Carlo (MCMC) sampling procedure, specifically utilizing the Metropolis-Hastings (MH) algorithm.

The MWR algorithm operates in an iterative two-step process:

  1. Proposal Step: A noisy version of the current image is created by adding Gaussian noise. Then, a projection operator is applied to enforce that the known Fourier coefficients (in the "sampled region") are preserved, while the "missing wedge" region is filled with noisy Fourier coefficients. This is followed by a denoising step using a chosen algorithm (e.g., BM3D, BM4D) to remove the added noise and regularize the image. This generates a candidate image.
  2. Evaluation Step: The candidate image is compared to the current image using a cost function (energy). Based on this comparison and the Metropolis-Hastings acceptance probability, the candidate image is either accepted as the new current image or rejected, keeping the previous one.

This iterative process allows the algorithm to explore the space of possible solutions and converge towards a statistically optimal estimate. A key innovation is the proposal mechanism, which involves perturbation, projection, and denoising, designed to efficiently explore the solution space and achieve good acceptance rates.

Datasets

The paper evaluates its method on:

  • Synthetic Data: A 2D Shepp-Logan phantom (Fig. 1) and a 3D dataset (Dataset A) of a 20S proteasome corrupted by additive white Gaussian noise and an artificial missing wedge. This dataset was used to test various parameters and compare against other methods.
  • Real Cryo-ET Data:
    • Dataset B: A sub-tomogram of a gold particle.
  • Dataset C: A sub-tomogram of 80S ribosomes attached to a membrane.
  • Dataset D: A sub-tomogram of E. coli bacteria, including both single-axis and double-axis acquired data for comparison.

Key Insights and Ideas

  • MMSE as a principled approach: Framing the problem as MMSE estimation provides a statistically sound basis for reconstructing the missing information, going beyond simple interpolation or filtering.
  • MCMC for high-dimensional integrals: The paper demonstrates the effectiveness of MCMC, specifically the MH algorithm, as a practical tool for computing MMSE estimators in complex, high-dimensional inverse problems like cryo-ET reconstruction.
  • Novel Proposal Mechanism: The proposed three-step proposal (perturbation, projection, denoising) is a key algorithmic contribution, designed to balance exploration of the solution space with efficient convergence.
  • Flexibility in Denoising: The MWR framework is agnostic to the specific denoising algorithm used, allowing integration with state-of-the-art denoisers like BM3D/BM4D.
  • Improved Visualisation and Analysis: The restored images not only look better but also facilitate downstream analysis tasks like segmentation and classification by reducing artifacts.

Key Results

  • Significant artifact reduction: The MWR method demonstrably reduces ray artifacts, side artifacts, and object elongation caused by the missing wedge.
  • Noise reduction: The method effectively denoises cryo-ET images.
  • Quantitative improvements: On synthetic data, MWR achieved competitive or superior PSNR and CCC (Constrained Correlation Coefficient) values compared to other methods. For example, on Dataset A, MWR achieved a PSNR of 27.0 dB for MAP estimation and 27.9 dB for MMSE estimation, outperforming some competing methods.
  • Faster convergence: Compared to a similar approach by Maggioni et al. (2013), MWR showed faster convergence.
  • Real-world applicability: Experiments on real cryo-ET data (gold particle, ribosomes, bacteria) showed improved image quality, enhanced contrast, and better object sphericity/shape recovery. For Dataset D, MWR achieved a CCC score of 0.71, significantly higher than BM4D's 0.29-0.39, indicating better recovery of Fourier coefficients compared to the double-axis ground truth.

Limitations and Future Work

  • Computational Cost: While faster than some iterative methods, MWR is still computationally intensive, with a single-axis 64x64x64 volume taking approximately 5.5 minutes on a standard laptop (unoptimized Matlab version).
  • Parameter Sensitivity: The performance is influenced by parameters like noise variance and the scaling parameter β, though the authors suggest that some parameters do not require precise tuning.
  • Future Work: The authors suggest parallelizing the Markov chains to further speed up computation and exploring GPU implementation for processing larger volumes. They also note that while the method improves visualization and analysis, it doesn't recover truly unobserved data but rather makes the "best statistical guess."

Summary updated

Abstract

We propose a statistical method to address an important issue in cryo-electron tomography image analysis: reduction of a high amount of noise and artifacts due to the presence of a missing wedge (MW) in the spectral domain. The method takes as an input a 3D tomogram derived from limited-angle tomography, and gives as an output a 3D denoised and artifact compensated volume. The artifact compensation is achieved by filling up the MW with meaningful information. To address this inverse problem, we compute a Minimum Mean Square Error (MMSE) estimator of the uncorrupted image. The underlying high-dimensional integral is computed by applying a dedicated Markov Chain Monte-Carlo (MCMC) sampling procedure based on the Metropolis-Hasting (MH) algorithm. The proposed MWR (Missing Wedge Restoration) algorithm can be used to enhance visualization or as a pre-processing step for image analysis, including segmentation and classification of macromolecules. Results are presented for both synthetic data and real 3D cryo-electron images.
BibTeX
@article{moebel2019monte,
  author = {Emmanuel Moebel and Charles Kervrann},
  title = {{A Monte Carlo framework for missing wedge restoration and noise removal in cryo-electron tomography}},
  journal = {Journal of Structural Biology X},
  year = {2019},
  doi = {10.1016/j.yjsbx.2019.100013},
}
Figure