Motion detection in diffraction tomography by common circle methods

Michael Quellmalz, Peter Elbau, Otmar Scherzer, Gabriele Steidl

TL;DR

This paper introduces a novel "common circle" framework to determine the unknown rotation and translation of an object during optical diffraction tomography. By exploiting the fact that measurements in diffraction tomography correspond to hemispheres in Fourier space, the authors show that rigid motion parameters can be recovered by matching data along the circular arcs where these hemispheres intersect.

What Problem Does This Paper Address?

In many imaging modalities, like X-ray Computed Tomography (CT), we assume light travels in straight lines. However, when imaging tiny objects like biological cells, light diffracts, and we must account for its wave-like nature using Optical Diffraction Tomography (ODT).

A major challenge in ODT occurs when the object moves (rotates or translates) unpredictably during the scan—for instance, when a cell is suspended in "optical tweezers." If we don't know the exact orientation of the object for each frame, we cannot reconstruct its internal structure. While the "method of common lines" is a standard solution for this in Cryo-Electron Microscopy (cryo-EM), it relies on the Fourier Slice Theorem, which assumes straight-line projections. Because ODT follows the Fourier Diffraction Theorem, the data maps to curved surfaces (hemispheres) rather than flat planes, making the traditional "common lines" approach inapplicable.

Methods

The authors propose two primary strategies based on the Fourier Diffraction Theorem:

  1. The Direct Common Circle Method: Instead of planes intersecting in lines (as in cryo-EM), ODT measurements represent hemispheres in 3D Fourier space. Two different hemispheres intersect in a circular arc. By comparing the values of the measured data along these arcs, the algorithm searches for the relative Euler angles that minimize the difference between the two datasets.
  2. The Infinitesimal Common Circle Method: If the object's motion is smooth and continuous, the authors derive a method using first-order derivatives (in both time and space). This allows them to calculate the angular velocity of the object at any given moment, which can then be integrated to find the full rotation path.
  3. Translation Recovery: Unlike standard X-ray projections, which are invariant to translations along the direction of the beam, ODT data is sensitive to the object's position. The authors demonstrate that if the scattering potential is real-valued, the full 3D translation vector can be recovered using phase information from the common circles.

The paper also highlights a mathematical "bridge": using stereographic projection, these common circles can be mapped back into straight lines, allowing ODT to be treated with tools similar to those used in classical tomography.

Datasets

The study uses "in silico" (simulated) datasets to provide a numerical proof-of-concept:

  • Cell Phantom: A simulated biological cell consisting of various convex and concave shapes with constant values.
  • Shepp-Logan Phantom: A standard 3D test object used in imaging research to represent head-like structures.
  • Simulated Motion: The authors applied both constant rotation axes and time-dependent "moving" rotation axes, along with complex 3D translations, to test the robustness of the algorithms.

Key Insights and Ideas

  • Two is Enough: In classical cryo-EM, you generally need to compare three images to uniquely determine relative orientations. In ODT, because the data lies on hemispheres, two images are sufficient to find the relative rotation.
  • Translation is Visible: A surprising "aha" moment is that ODT is sensitive to the third component of translation (depth). In standard ray-transform imaging, moving an object closer to or further from the source doesn't change the projection; in ODT, it changes the phase, allowing for full 3D motion tracking.
  • Hybrid Approach: The infinitesimal method is computationally very fast but tends to accumulate "drift" errors over time. The authors found that using the infinitesimal method to provide a "starting guess" for the direct common circle optimization yields the best balance of speed and high precision.

Key Results

  • High Precision: In simulations with 640 time steps, the relative error for the reconstructed rotation matrices was remarkably low, often in the range of $10^{-3}$ to $10^{-5}$.
  • Robustness: The method successfully handled complex motions where the rotation axis itself changed over time.
  • Reconstruction Quality: Using the recovered motion parameters, the authors were able to reconstruct the 3D phantoms with high fidelity, achieving a Peak Signal-to-Noise Ratio (PSNR) of over 30 dB and high Structural Similarity Index (SSIM) values (~0.75).
  • Efficiency: Reconstructing motion parameters for 640 time steps took approximately 40 seconds on a standard consumer PC.

Limitations and Future Work

  • Born Approximation: The current math relies on the "Born approximation," which assumes the object scatters light weakly. This is usually true for small biological cells but fails for more opaque or highly refractive objects.
  • Real-Valued Potential: The translation recovery currently assumes the object does not absorb light (real-valued scattering potential). Extending this to complex-valued potentials (objects that both scatter and absorb) is a future goal.
  • Phase Retrieval: Real-world detectors often only measure light intensity, losing the "phase" information. The authors note that incorporating phase retrieval methods will be essential for applying this to experimental laboratory data.
  • Experimental Validation: The next step is applying these methods to real-world ODT data involving cells trapped in acoustical or optical tweezers.

Summary updated

Abstract

: The method of common lines is a well-established reconstruction technique in cryogenic electron microscopy (cryo-EM), which can be used to extract the relative orientations of an object given tomographic projection images from different directions. In this paper, we deal with an analogous problem in optical diffraction tomography. Based on the Fourier diffraction theorem, we show that rigid motions of the object, i.e., rotations and translations, can be determined by detecting common circles in the Fourier-transformed data. We introduce two methods to identify common circles. The first one is motivated by the common line approach for projection images and detects the relative orientation by parameterizing the common circles in the two images. The second one assumes a smooth motion over time and calculates the angular velocity of the rotational motion via an infinitesimal version of the common circle method. Interestingly, using the stereographic projection, both methods can be reformulated as common line methods, but these lines are, in contrast to those used in cryo-EM, not confined to pass through the origin and allow for a full reconstruction of the relative orientations. Numerical proof-of-the-concept examples demonstrate the performance of our reconstruction methods.
BibTeX
@inproceedings{quellmalz2023motion,
  author = {Michael Quellmalz and Peter Elbau and Otmar Scherzer and Gabriele Steidl},
  title = {{Motion detection in diffraction tomography by common circle methods}},
  booktitle = {Math. Comput},
  year = {2023},
  url = {https://openreview.net/forum?id=hVCCGN0Hri},
}

Loading related content...

Figure