Sub-Resolution Axial Localization of Nanoparticles in Fluorescence Microscopy

François Aguet, Dimitri Van De Ville, Michaël Unser2005 · DOI

TL;DR

This paper presents a novel method for precisely determining the 3D position of fluorescent nanoparticles within a microscope, even when they are out of focus. By employing a maximum-likelihood estimation technique with an accurate model of how light behaves in a microscope, they achieve nanometer-level precision in axial (z-axis) localization, surpassing previous empirical methods.

What Problem Does This Paper Address?

Understanding the dynamics of individual molecules within living cells is crucial for biological research. Fluorescence microscopy allows us to visualize these molecules, but accurately tracking their movement in three dimensions (x, y, and z) is challenging. While lateral (x-y) localization has seen significant advancements, precisely determining a particle's axial position (depth) has remained a bottleneck. This is largely due to the complex way light diffracts and forms an image (the point spread function or PSF), which changes significantly as the particle moves away from the focal plane. Existing methods often rely on simplified models or empirical approaches that lack accuracy.

Methods

The core of this paper is a maximum-likelihood estimation (MLE) algorithm designed to pinpoint a particle's axial position. The authors start by developing a realistic image formation model that accounts for the non-stationary nature of the PSF – meaning how the PSF changes with depth. This model treats the observed photon counts in each pixel of the CCD camera as following a Poisson distribution, a standard assumption for photon noise.

The key innovation lies in how they use this model:

  • Accurate PSF Model: Instead of simplified approximations, they utilize a more accurate PSF model that considers factors like refractive index differences between the specimen and immersion medium.
  • Cramér-Rao Bound (CRB): They derive the theoretical lower bound on the precision achievable for axial localization, providing a benchmark for their method.
  • Iterative MLE: Since directly solving for the axial position (zₚ) from the complex PSF model is difficult, they linearize the model using a first-order Taylor approximation. This allows them to derive an iterative update formula for zₚ, refining the estimate with each step.
  • Initial Estimate: To kickstart the iterative process, they propose using normalized cross-correlation between the acquired image and pre-computed PSF sections.

Datasets

Not discussed in the paper. The authors mention generating results using a theoretical PSF model and simulating acquisitions.

Key Insights and Ideas

  • Defocus is Your Friend (for Axial Localization): Counterintuitively, imaging particles off-focus can lead to greater accuracy in axial position estimation. This is because the diffraction pattern changes more dramatically with depth when out of focus, providing more information for the algorithm to exploit.
  • The Power of a Good Model: The accuracy of the axial localization is highly dependent on the accuracy of the PSF model used. A model that correctly accounts for the changing PSF with depth is essential.
  • Theoretical Limits are Achievable: The proposed MLE method achieves axial localization precision that closely matches the theoretical Cramér-Rao bound, suggesting it's a near-optimal approach.

Key Results

The paper demonstrates that their MLE method can achieve nanometer-range precision in axial localization. Figure 1(b) shows that the standard deviation of their estimation closely matches the predicted Cramér-Rao bound across various defocus distances. They also show that with multiple acquisitions, their method can help determine the optimal experimental configuration (e.g., focal plane positions for each acquisition) to maximize precision over a specific region of interest.

Limitations and Future Work

  • Initial Estimate Sensitivity: While they propose a cross-correlation method for an initial estimate, the paper doesn't extensively detail its robustness or how sensitive the final result is to this initial guess.
  • Computational Cost: While not explicitly stated as a limitation, iterative MLE methods can be computationally intensive, especially with many acquisitions or high-resolution images.
  • Idealized Conditions: The results are presented under what appear to be relatively ideal conditions (e.g., single particles, well-defined PSFs). Real biological samples can be much more complex.

Future work could involve extending this method to handle multiple particles simultaneously, exploring its performance in more complex biological environments, and optimizing the computational efficiency for real-time applications.

Summary updated

Abstract

We propose a method for sub-resolution axial localization of particles in fluorescence microscopy, based on maximum-likelihood estimation. Given acquisitions of a defocused fluorescent particle, we are can estimate its axial position within nanometer range precision.
BibTeX
@misc{aguet2005sub,
  author = {François Aguet and Dimitri Van De Ville and Michaël Unser},
  title = {{Sub-Resolution Axial Localization of Nanoparticles in Fluorescence Microscopy}},
  year = {2005},
  doi = {10.1364/ecbo.2005.pdb1},
}

Loading related content...

Figure