Spline-based image-to-volume registration for three-dimensional electron microscopy

Slavica Jonić, Carlos Óscar S. Sorzano, P. Thévenaz, C. El-Bez, Sacha De Carlo, Michaël UnserUltramicroscopy 2005 · DOI

TL;DR

This paper introduces a novel frequency-domain algorithm for precisely aligning 3D electron microscopy (EM) images to a reference 3D model. By leveraging B-spline interpolation and a continuous parameterization, their method offers improved accuracy and robustness compared to existing techniques, crucial for reconstructing high-resolution biological structures.

What Problem Does This Paper Address?

Three-dimensional electron microscopy (3DEM) is a powerful technique for visualizing biological structures at the molecular level. A key challenge in 3DEM is accurately reconstructing a 3D model from a collection of 2D projection images. This reconstruction process typically involves two main steps: first, determining the precise orientation and position (angular and translational assignment) of each 2D image relative to the 3D structure, and second, using these assignments to build the 3D model.

Existing methods for angular and translational assignment often rely on quantized parameters, meaning they discretize the possible orientations and positions. This can lead to a loss of accuracy, as the true optimal alignment might fall between these discrete steps. While some methods use continuous parameters, they can be sensitive to initial guesses and computationally expensive. This paper aims to address these limitations by developing a more accurate, robust, and efficient continuous-parameter registration algorithm in the frequency domain.

Methods

The core of the proposed method is a frequency-domain, continuous-parameter image-to-volume registration algorithm. Here's a breakdown of the key techniques:

  • Fourier Transform (FT) Domain Registration: Instead of working directly with image pixel intensities, the algorithm operates on the Fourier transforms of the EM images and the projections of the 3D reference volume. This is motivated by the fact that the Fourier transform often reveals structural information more directly and can be computationally advantageous.
  • Central Slice Theorem (CST): The CST is a fundamental principle that relates the 2D FT of a projection of a 3D object to a slice through the 3D FT of that object. The algorithm uses the CST to efficiently extract the relevant slice from the 3D FT of the reference volume that corresponds to the orientation of the 2D EM image.
  • Cubic B-spline Interpolation: To accurately extract these slices from the discrete 3D FT of the volume and to compute the necessary gradients for optimization, the algorithm models the 3D FT using a continuous representation based on cubic B-splines. This provides a smooth and well-defined mathematical framework, particularly for calculating derivatives.
  • Gradient-Based Iterative Minimization: The algorithm employs a Levenberg-Marquardt (LM) optimizer to iteratively minimize a least-squares measure of dissimilarity between the 2D FT of the EM image and the corresponding slice from the 3D FT of the volume. The LM algorithm is chosen for its balance between robustness and efficiency, especially when gradients are available.
  • "Mixed" Initialization Strategy: To improve robustness against poor initial guesses, the algorithm uses a novel strategy. It launches the optimization process from multiple initial assignments (obtained from standard quantized-parameter methods like MPM, PM, and RA) and then selects the assignment that results in the minimum final dissimilarity measure.
  • Weighted Cost Function: The dissimilarity measure is weighted to attenuate high frequencies and discard the DC component. This acts as a band-pass filter, improving robustness by down-weighting noisy high-frequency information.

Datasets

The paper uses two types of datasets for validation:

  • Synthesized EM Images: The authors generated synthetic cryo-EM images from known 3D protein models (Bacteriorhodopsin and Hexon from the Protein Data Bank - PDB). These synthetic images were created by projecting the 3D models with known orientations and translations, and then adding simulated noise and the effect of the Contrast Transfer Function (CTF). This allows for objective evaluation of the algorithm's accuracy against a "ground truth."
  • Real EM Data: The algorithm was also tested on experimental cryo-EM images of the GroEL chaperonin, where the ground truth alignment was unknown. This dataset was used to evaluate the algorithm's ability to refine existing 3D models and improve the consistency of reconstructions.

Key Insights and Ideas

  • Frequency Domain Advantage: Operating in the Fourier domain, particularly with B-spline interpolation, allows for more accurate and robust estimation of alignment parameters compared to space-domain methods. This is attributed to better satisfaction of the space-limitedness assumption in the frequency domain.
  • Continuous Parameters are Key: Moving from quantized to continuous parameters for orientation and translation significantly improves registration accuracy, avoiding the limitations of discrete steps.
  • B-splines for Smoothness and Gradients: The use of cubic B-splines is crucial for creating a continuous model of the 3D FT, enabling accurate slice extraction and, importantly, the computation of well-defined gradients for the optimization process.
  • "Mixed" Strategy for Robustness: The novel initialization strategy, combining multiple starting points from established methods, dramatically improves the algorithm's robustness, making it less sensitive to the quality of the initial guess.
  • Low-Resolution Models Can Be Beneficial: Surprisingly, the paper shows that using a lower-resolution reference model can sometimes lead to better convergence and smaller standard deviations in misregistration, suggesting that noise and high-frequency artifacts in high-resolution models can hinder optimization.

Key Results

  • Improved Accuracy: The algorithm achieved an assignment accuracy better than 2 voxels (approximately 6.54 Å) for both synthesized and real data when using the "mixed" initialization strategy.
  • Robustness: The frequency-based method demonstrated superior robustness compared to a space-based registration method, especially when dealing with larger initial misalignments.
  • Efficiency: While not explicitly quantified in terms of speedup over all methods, the use of the LM optimizer and the frequency-domain approach is implied to be efficient. For example, it achieved a better 0.5-FSC resolution in fewer reconstructions than a standard PM-based refinement.
  • Consistency: On real EM data (GroEL), the algorithm significantly improved the consistency of 3D reconstructions, achieving a better 0.5-FSC resolution (12.5 Å) in fewer iterations than a PM-based approach.

Limitations and Future Work

  • Initialization Sensitivity (Mitigated): While the "mixed" strategy significantly improves robustness, the algorithm still requires initial guesses. The paper demonstrates that even a coarser library for initial assignment (10° step vs. 5°) does not lead to a significant loss in accuracy, suggesting flexibility.
  • Data Assumptions: The accuracy of the frequency-domain interpolation relies on the assumption that the data are "nearly space-limited." This is addressed by spatial apodization with Gaussian windows, but it's an underlying assumption.
  • Future Work: The paper suggests that the "mixed" strategy can be generalized to include more initial assignment methods. The availability of the algorithm in the Xmipp package [43] indicates a path for broader adoption and further development. The authors also highlight the potential for this method to be combined with various reconstruction algorithms for iterative refinement.

Summary updated

BibTeX
@inproceedings{jonic2005spline,
  author = {Slavica Jonić and Carlos Óscar S. Sorzano and P. Thévenaz and C. El-Bez and Sacha De Carlo and Michaël Unser},
  title = {{Spline-based image-to-volume registration for three-dimensional electron microscopy}},
  booktitle = {Ultramicroscopy},
  year = {2005},
  doi = {10.1016/j.ultramic.2005.02.002},
}
Figure