Practical sensorless aberration estimation for 3D microscopy with deep learning.

Area of Science:

• Optical microscopy
• Computational imaging
• Machine learning for science

Background:

• Accurate estimation of optical aberrations is crucial for sensorless adaptive optics in 3D microscopy.
• Deep learning methods offer potential for high-speed aberration correction but require extensive ground truth data for training.
• Acquiring ground truth microscopy data is often impractical, hindering the application of data-driven approaches.

Purpose of the Study:

• To demonstrate that neural networks trained solely on simulated data can accurately predict optical aberrations in real experimental microscopy images.
• To validate the simulation-based training approach across different microscopy modalities and compare it with non-learned methods.
• To investigate the data requirements for predicting individual aberrations, highlighting the role of wavefront symmetry.

Main Methods:

• Development and training of neural networks using exclusively simulated volumetric intensity images.
• Validation of the trained networks on both simulated and experimentally acquired datasets from two distinct microscopy modalities.
• Comparative analysis against established non-learned methods for aberration estimation.
• Analysis of aberration predictability based on data requirements and wavefront symmetry properties.

Main Results:

• Neural networks trained on simulated data achieved accurate predictions for optical aberrations in real experimental images.
• The approach demonstrated effectiveness across different microscopy modalities.
• Wavefront symmetry was identified as a critical factor influencing the predictability of individual aberrations.
• Performance was comparable or superior to non-learned methods in certain aspects.

Conclusions:

• Simulated data is sufficient for training deep learning models for optical aberration estimation in 3D microscopy.
• This overcomes the practical limitations of ground truth data acquisition, making sensorless adaptive optics more accessible.
• The open-source availability of the Python implementation facilitates further research and application.