Convolutional Neural Networks for Phase Prediction in Deep Tissue Microscopy

Adaptive optics have the potential to improve the resolution of optical imaging in deep tissue up to the diffraction limit. We study neural networks as a possible tool to help improve the convergence rate of adaptive optics systems by predicting the wavefront phase error from an abberated PSF.
Adaptive optics systems have proved successful in correcting aberrations produced by atmospheric conditions in astronomy imaging. In this setting artificial guide stars and wavefront sensors which can be used to reconstruct spatial phase variation in the signal wavefront. In deep tissue, given a comparable guide star “particle”, wavefront sensors fail to accurately measure phase error due to the highly scattering nature of the tissue. Thus, alternative strategies to optimize the array positions of the deformable mirror are necessary.
It has previously been shown that a stochastic gradient descent method which randomly permutes the degrees of freedom of the mirror can optimize to an improvement of 2-5x peak intensity. However this process is time consuming (necessitating between 600 and 3000 camera integrations to find an optimal array position). We would like to train a neural network to predict phase distortions directly from distorted PSF images.
In this thesis we show a proof of concept by predicting Zernike and Kolmogorov phase screens from corresponding PSF images. We further show that convolutional neural network architectures appear superior to fully-connected neural network architectures in phase prediction. Finally, we demonstrate a system for the mass collection of data to train a neural network based on real biological tissue. This study represents the first research conducted to use neural networks for adaptive optics for microscopy, and the first research to use convolutional neural networks for any type of adaptive optics.