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

Emmanuel Moebel, Charles KervrannHAL 2018

TL;DR

This paper introduces a novel Monte Carlo framework using a Metropolis-Hastings algorithm to restore missing information in cryo-electron tomography (cryo-ET) images caused by a "missing wedge" artifact. This method effectively denoises images and compensates for artifacts, leading to improved visualization and analysis of biological structures.

What Problem Does This Paper Address?

Cryo-electron tomography (cryo-ET) is a powerful technique for visualizing cellular structures at near-atomic resolution. However, two major challenges limit its effectiveness: 1) high noise levels due to low electron doses required to preserve sample integrity, and 2) artifacts arising from the "missing wedge" (MW) in the spectral domain. This MW is a consequence of the limited tilting range of the electron microscope, meaning some angular projections are unavailable. This missing information leads to artifacts like Gibbs phenomena (ray and side artifacts) and object elongation, obscuring fine details and hindering accurate analysis of molecular complexes. Existing methods often struggle to fully address both noise and the MW artifact simultaneously, especially in complex biological samples.

Methods

The core of this work is a statistical approach based on Bayesian inference and Monte Carlo Markov Chain (MCMC) sampling, specifically using the Metropolis-Hastings (MH) algorithm. The authors formulate the problem as recovering a clean image ($x$) from a noisy, artifact-ridden one ($y$) by estimating the Minimum Mean Square Error (MMSE) of the uncorrupted image.

Key methodological contributions include:

  • MMSE Estimator for MW Restoration: They propose a specific MMSE estimator tailored for the MW restoration problem, incorporating a prior that encourages positive and piecewise smooth solutions. This prior is defined within a constrained space that enforces similarity to known Fourier coefficients and limits total variation.
  • Original MCMC Sampling Procedure: To compute the high-dimensional integral required for the MMSE estimator, they develop a novel MH algorithm. This algorithm involves a three-step proposal mechanism:
  • Perturbation: Adding Gaussian noise to the current estimate.
  • Projection: Enforcing known Fourier coefficients within the sampled region (SR).
  • Denoising: Applying a denoising operation (e.g., Total Variation) to ensure smoothness and reduce artifacts.
  • Adaptive Proposal Density: The MH algorithm's proposal density is designed to be approximately symmetric and aims to explore the solution space efficiently, balancing acceptance rate and sample correlation.
  • "Periodic plus smooth image decomposition": An additional step is incorporated at the end of the proposal stage to mitigate cross-structure artifacts that can arise from the Fourier transform.

Datasets

The paper evaluates its method on both synthetic and real experimental data:

  • Synthetic Data (Dataset A): A 3D density map of the 20S proteasome, corrupted by additive white Gaussian noise and an artificial MW process. This dataset is used to optimize parameters, assess robustness to noise, and compare with other methods.
  • Real Experimental Data:
    • Dataset B: An experimental sub-tomogram of a gold particle.
  • Dataset C: An experimental sub-tomogram of 80S ribosomes attached to a membrane.
  • Dataset D: An experimental sub-tomogram of an E. coli bacteria region, available in both single-axis and double-axis versions. The double-axis data serves as a ground truth for the single-axis data.

Key Insights and Ideas

  • MMSE over MAP for Noise: The MMSE estimator, while visually similar to MAP, produces less background noise and a higher PSNR, indicating it's closer to the ground truth. This is because MAP can be more sensitive to noise in high-frequency components.
  • Importance of Data Fidelity Term: While various data fidelity terms (L1, L2, correlation, PSNR, MI) were tested, the resulting images were visually similar, suggesting the method is robust to the specific choice of this term. The L2 norm was chosen for consistency.
  • Fourier Transform is Key: The discrete fast Fourier transform (FFT) proved to be the most effective spectral transform for MW restoration, closely followed by the discrete cosine transform (DCT). Other transforms like wavelet were less suitable.
  • MCMC Efficiency: The proposed MH algorithm offers a faster convergence than previous MCMC-based methods for MW restoration, achieving comparable or better results.
  • Handling Real-World Noise: The method demonstrates effectiveness even with significant noise levels in cryo-ET data, though performance is naturally better with higher signal-to-noise ratios.

Key Results

  • Improved Image Quality: The method significantly reduces noise and MW artifacts, enhancing image contrast and detail.
  • Quantitative Improvements:
    • On synthetic data (Dataset A), the method achieved higher PSNR and CCC scores compared to simple denoising and other competing methods. For example, PSNR improved from 14.4 dB (corrupted) to 27.8 dB with the Fourier transform.
  • On real data (Dataset B), sphericity of a gold particle improved.
  • On Dataset C, signal enhancement was observed, and ribosomes became clearly visible.
  • On Dataset D, the CCC score rose from 0.29 to 0.44 when processing single-axis data against a double-axis ground truth.
  • Faster Convergence: The proposed MCMC approach converges approximately twice as fast as the BM4D method from Maggioni et al. (2013).
  • Comparison to Other Methods: The proposed method outperformed SMAPEM, BFLY, and Moisan's TV method in terms of PSNR and CCC scores, producing smoother images without the pixelated artifacts seen in some other approaches.

Limitations and Future Work

  • Contrast Dependency: The method requires a reasonable contrast of the object of interest to perform optimally. This is not always the case in cryo-ET, especially for low-SNR samples.
  • Parameter Tuning: While the method is shown to be robust to parameter choices (e.g., acceptance rate, data fidelity terms), some parameters still require careful consideration.
  • Future Directions: The authors suggest that with advancements in electron microscopy (e.g., direct electron detection, phase contrast methods), the method's effectiveness will further increase. They also propose that the framework can be applied to other imaging modalities with similar inverse problems, particularly those with high SNR values.

Summary updated

Abstract

In this paper, we describe 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 tomo-gram 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 computational method 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
@inproceedings{moebel2018monte,
  author = {Emmanuel Moebel and Charles Kervrann},
  title = {{A Monte Carlo framework for noise removal and missing wedge restoration in cryo-electron tomography}},
  booktitle = {HAL},
  year = {2018},
}
Figure