Skip to main content
Back to timeline
Computational and Structural Biotechnology JournalSource publication:

Dense fluorescence localization stops giving one answer: a diffusion model samples uncertainty in molecule count and position

Synopsis

In densely labeled single-molecule localization microscopy, one low-resolution image can be consistent with several molecular arrangements, while conventional deep models regress a single deterministic solution. A conditional variational diffusion model (CVDM) instead learns a distribution over high-resolution kernel density estimates, localizes multiple emitters jointly across the full field of view, and samples several plausible reconstructions from the same input. On simulated images, localization precision and recall decay as emitter density rises but remain strong over the tested density range, and reconstructions on DNA origami rulers and microtubule data agree with the known structures.

Source-provided article image: Uncertainty-Aware Localization Microscopy by Variational Diffusion.
PubMed

Interpretation

Deep models for dense single-molecule localization microscopy can now return several plausible solutions rather than one deterministic reconstruction. Earlier deep localization models output a single kernel density estimate or emitter occupancy probability, and Bayesian multi-emitter fitting gives a posterior only by explicitly modeling emitter number and coordinates. On 64×64 low-resolution simulated images, predicted kernel density estimates were compared with ground truth; localization error varies with emitter density, and precision and recall decay with increasing density while remaining strong over the tested density range.

The model treats localization as image restoration and learns a conditional distribution from low-resolution images to high-resolution kernel density estimates. The conditional variational diffusion model (CVDM) is a diffusion generative model that progressively destroys structure in the kernel density map and learns a reverse denoising process, generating plausible kernel density estimates given an observed image. Training covered a variety of simulated images while fixing the point spread function and noise characteristics of the experimental camera; on Alexa555-tagged DNA origami rulers spaced 94 nm apart, predicted localizations appeared in pairs at separations consistent with the known ruler spacing.

Multiple reconstructions can be sampled from the same input, and variation among samples exposes ambiguity in emitter number and position. Deterministic regression gives one point estimate and cannot express cases where several molecular arrangements fit the observation in dense scenes. Repeated sampling of the same image shows spots appearing and disappearing in the kernel density estimate, location variation correlates with localization error, and variation grows in higher-density samples.

Perspective

Validation covers zero-background simulated images, Alexa555-tagged 94 nm DNA origami rulers, and public high-density and sparse long-sequence microtubule data; the microtubule experiments were not newly acquired in this study. The intended readers are researchers working on superresolution microscopy and image restoration.

The zero-background simulations simplify the nonuniform background seen in experiments, and experimental images lacked suitable ground truth, so training relies on simulated approximations. The model mainly captures aleatoric uncertainty inherent to the image restoration problem and does not provide a directly interpretable posterior over emitter positions and number in coordinate space. Reconstructions can show low-level intensity variation around predicted spots and artifacts near image boundaries. Localization uncertainty is not guaranteed to scale as the inverse square root of photon count.