Markerauto2: A fast and robust fully automatic fiducial marker-based tilt series alignment software for electron tomography

Zihe Xu, Hongjia Li, Xiaohua Wan, Jose-Jesus Fernández, Fei Sun, Fa Zhang, Renmin HanStructure 2024 · OpenReview

Introduction

Cryo-Electron Tomography (cryo-ET) is a powerful technique for visualizing biological structures in 3D. However, the alignment of 2D tilt series images remains a significant computational bottleneck. Imperfections in the microscope stage and sample deformation require precise alignment, typically achieved using fiducial markers like gold beads. Tracking these markers across a series of images—often 60 or more—is challenging, particularly at high tilt angles where increased sample thickness leads to extremely low signal-to-noise ratios (SNR). Markerauto2 addresses these challenges by introducing a fully automated, high-speed, and robust alignment pipeline.

Methodology

Markerauto2 implements a four-stage pipeline designed to automate and accelerate the alignment process while maintaining high accuracy in noisy conditions.

The Markerauto2 alignment pipeline flowchart, illustrating the iterative loop of detection, matching, and calibration.

1. Wavelet-Based Detection

To overcome the limitations of simple template matching in low-SNR environments, Markerauto2 employs a wavelet multiscale template for initial marker detection. This approach is more resilient to noise and is parallelized using OpenMP, achieving a 5–20x speedup over traditional methods.

2. Local Geometric Hashing

Marker tracking is optimized using local geometric hashing based on 4-point affine invariant features. By transforming the matching problem into a hashing search, the computational complexity is reduced from $O(N^2 \log N)$ to near-linear $O(K \log K)$. This allows the software to process datasets in seconds that previously required hours.

3. Iterative Supplementation

Markers frequently "disappear" at high tilt angles due to noise. Markerauto2 uses 3D coordinates derived from low-angle images to predict marker locations at high angles. This iterative process fills gaps in tracking chains, ensuring a more complete dataset for reconstruction.

The iterative supplementation process: 3D coordinates from low-tilt images are reprojected to predict and recover marker positions in high-tilt frames.

4. Group-Weighted Calibration

Standard alignment algorithms often penalize high-tilt markers due to their naturally higher residuals. Markerauto2 introduces a group-weighted strategy using Median Absolute Deviation (MAD). By grouping weights by tilt angle, the algorithm avoids discarding vital high-angle data, recognizing that increased noise is a systematic property of the physics rather than a marker error.

Results and Discussion

The software was evaluated using 10 real-world datasets (including HIV-1 Gag and SARS-VLP) and simulated data with high noise and outliers.

Accuracy and Robustness

Markerauto2 consistently achieved lower reprojection residuals compared to industry standards. For the HIV-1 Gag dataset, it reached a residual of 0.65 pixels, significantly outperforming IMOD's 2.9 pixels. It successfully aligned complex datasets where previous tools failed, demonstrating superior robustness in low-SNR and high-marker-density scenarios.

Computational Efficiency

The most significant improvement is in processing speed. For the "Nitrosop3" dataset, tracking time was reduced from 13.4 hours in the original Markerauto to just 12.9 seconds in Markerauto2—a speedup of approximately 3,700x.

Conclusion

Markerauto2 represents a major advancement in cryo-ET data processing, transforming a manual bottleneck into a rapid, automated step. While it remains dependent on fiducial markers, its robust weighting and iterative refinement strategies set a new standard for alignment software. Future work may extend these robust estimation techniques to other stages of the cryo-ET workflow, such as Contrast Transfer Function (CTF) estimation.

Summary updated

BibTeX
@inproceedings{xu2024markerauto,
  author = {Zihe Xu and Hongjia Li and Xiaohua Wan and Jose-Jesus Fernández and Fei Sun and Fa Zhang and Renmin Han},
  title = {{Markerauto2: A fast and robust fully automatic fiducial marker-based tilt series alignment software for electron tomography}},
  booktitle = {Structure},
  year = {2024},
  url = {https://openreview.net/forum?id=VDyaB4tXOL},
}
Figure