Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

J-Invariant Volume Shuffle for Self-Supervised Cryo-Electron Tomogram Denoising on Single Noisy Volume

Xiwei Liu, Mohamad Kassab, Min Xu, Qirong Ho

TL;DR

This work tackles the challenge of denoising Cryo-ET volumes without paired data by introducing a J-invariant blind-spot U-shaped network that denoises a single noisy volume. Core innovations include sparse centrally masked convolution to preserve $\,\mathcal{J}$-invariance, volume-unshuffle/shuffle to expand receptive fields while maintaining invariance, and dilated channel attention with an edge-guided loss framework. The method employs a multi-term loss balancing structural fidelity and smoothness, and demonstrates superior performance over existing self-supervised approaches on simulated and real Cryo-ET datasets, with robust preservation of ultrastructural details. Overall, the approach advances self-supervised tomographic denoising, enabling more accurate visualization of cellular architecture in structural biology.

Abstract

Cryo-Electron Tomography (Cryo-ET) enables detailed 3D visualization of cellular structures in near-native states but suffers from low signal-to-noise ratio due to imaging constraints. Traditional denoising methods and supervised learning approaches often struggle with complex noise patterns and the lack of paired datasets. Self-supervised methods, which utilize noisy input itself as a target, have been studied; however, existing Cryo-ET self-supervised denoising methods face significant challenges due to losing information during training and the learned incomplete noise patterns. In this paper, we propose a novel self-supervised learning model that denoises Cryo-ET volumetric images using a single noisy volume. Our method features a U-shape J-invariant blind spot network with sparse centrally masked convolutions, dilated channel attention blocks, and volume unshuffle/shuffle technique. The volume-unshuffle/shuffle technique expands receptive fields and utilizes multi-scale representations, significantly improving noise reduction and structural preservation. Experimental results demonstrate that our approach achieves superior performance compared to existing methods, advancing Cryo-ET data processing for structural biology research

J-Invariant Volume Shuffle for Self-Supervised Cryo-Electron Tomogram Denoising on Single Noisy Volume

TL;DR

This work tackles the challenge of denoising Cryo-ET volumes without paired data by introducing a J-invariant blind-spot U-shaped network that denoises a single noisy volume. Core innovations include sparse centrally masked convolution to preserve -invariance, volume-unshuffle/shuffle to expand receptive fields while maintaining invariance, and dilated channel attention with an edge-guided loss framework. The method employs a multi-term loss balancing structural fidelity and smoothness, and demonstrates superior performance over existing self-supervised approaches on simulated and real Cryo-ET datasets, with robust preservation of ultrastructural details. Overall, the approach advances self-supervised tomographic denoising, enabling more accurate visualization of cellular architecture in structural biology.

Abstract

Cryo-Electron Tomography (Cryo-ET) enables detailed 3D visualization of cellular structures in near-native states but suffers from low signal-to-noise ratio due to imaging constraints. Traditional denoising methods and supervised learning approaches often struggle with complex noise patterns and the lack of paired datasets. Self-supervised methods, which utilize noisy input itself as a target, have been studied; however, existing Cryo-ET self-supervised denoising methods face significant challenges due to losing information during training and the learned incomplete noise patterns. In this paper, we propose a novel self-supervised learning model that denoises Cryo-ET volumetric images using a single noisy volume. Our method features a U-shape J-invariant blind spot network with sparse centrally masked convolutions, dilated channel attention blocks, and volume unshuffle/shuffle technique. The volume-unshuffle/shuffle technique expands receptive fields and utilizes multi-scale representations, significantly improving noise reduction and structural preservation. Experimental results demonstrate that our approach achieves superior performance compared to existing methods, advancing Cryo-ET data processing for structural biology research

Paper Structure

This paper contains 17 sections, 1 theorem, 11 equations, 7 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Preliminaries
  4. Methodology
  5. Network Architecture
  6. Main Components
  7. Loss Function
  8. Experiments
  9. Networks Training Details
  10. Experiments on Simulated Data
  11. Real-world Datasets
  12. Experiments on Real Cryo-ET data
  13. Ablation studies
  14. Component analysis of loss function.
  15. Study on noise patterns in simulated data.
  16. ...and 2 more sections

Key Result

Lemma 1

The additive Gaussian noise in 2D projection remains Gaussian noise in the 3D reconstruction.

Figures (7)

  • Figure 1: The architecture of proposed method. The noisy input volume is first preprocessed using (i) 3D Gaussian filters to generate a smoothed volume and (ii) bilateral filters to create a edge preserved filtered volume, respectively. We construct the U-shape BSN with 4 main contributions: (1) Sparse centrally masked convolution, (2) Volume Unshuffle/Shuffle and (3) DCA blocks, with (4) an edge representation enhancer utilizes the filtered volume to guide model training. Refer to Section \ref{['sec:main_components']} for details.
  • Figure 2: Volume unshuffle and shuffle.
  • Figure 3: Details of edge representation enhancer.
  • Figure 4: Visual results of the real data.
  • Figure 5: Examples of FSC$_{e/o}$ curves for the G.hanseni and Phage. Red dash line in the figures point out the position of FSC$_{e/o}$=0.5.
  • ...and 2 more figures

Theorems & Definitions (2)

  • Lemma 1
  • Definition 1