Learned, uncertainty-driven adaptive acquisition for photon-efficient scanning microscopy.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Machine Learning

Background:

• Scanning microscopy (confocal, multiphoton) enables deep tissue imaging but faces trade-offs between speed, field of view, phototoxicity, and image quality.
• Deep learning offers denoising but risks hallucinating artifacts, which is problematic for scientific and medical applications.

Purpose of the Study:

• To develop a trustworthy deep learning method for scanning microscopy that simultaneously denoises images and quantifies pixel-wise uncertainty.
• To introduce an adaptive acquisition strategy leveraging uncertainty maps to optimize imaging time and reduce light dose.

Main Methods:

• A novel deep learning approach was developed to perform simultaneous denoising and pixel-wise uncertainty prediction for microscopy data.
• An adaptive acquisition technique was proposed, utilizing learned uncertainty maps to guide rescanning of uncertain sample regions.
• The method was validated on experimental confocal and multiphoton microscopy data.

Main Results:

• Uncertainty maps accurately identified hallucinations in deep learning-based denoised images.
• The adaptive acquisition technique achieved up to a 16x reduction in acquisition time and total light dose.
• Fine sample features were successfully recovered, and hallucinations were reduced with the adaptive method.

Conclusions:

• This work presents the first demonstration of distribution-free uncertainty quantification for denoising with real experimental microscopy data.
• The proposed adaptive acquisition strategy based on reconstruction uncertainty offers significant improvements in imaging efficiency and sample safety.