Uncertainty-aware localization microscopy by variational diffusion.

Area of Science:

• Biophysics
• Computational Biology
• Microscopy

Background:

• Deep learning accelerates fluorescence microscopy, particularly for single-molecule localization microscopy (SMLM).
• Kernel density (KD) estimation in SMLM is crucial for super-resolution imaging but challenging due to multiple potential solutions for localizing molecules in dense images.

Purpose of the Study:

• To propose a novel generative modeling framework for KD estimation in SMLM.
• To address the inverse problem of molecule localization by modeling a probability distribution of solutions.
• To enable uncertainty quantification in KD estimates, a capability lacking in current deep learning models for SMLM.

Main Methods:

• Developed a conditional variational diffusion model (CVDM) for KD estimation in SMLM.
• Trained the CVDM to perform localization tasks on low-resolution measurements by modeling high-resolution KD estimates.
• Utilized the model to probe the distribution structure of KD estimates and express uncertainty.

Main Results:

• Achieved high-fidelity super-resolution imaging of densely labeled structures.
• Enabled accurate uncertainty estimation for regressed KD estimates.
• Demonstrated the model's effectiveness for image restoration in SMLM.

Conclusions:

• The proposed CVDM framework offers a robust approach for molecule localization in SMLM.
• This method enhances super-resolution imaging by providing uncertainty quantification.
• The findings have significant implications for advancing single-molecule and super-resolution microscopy techniques.