Practical aberration correction using deep transfer learning with limited experimental data

Adaptive optics is a technique for correcting aberrations and improving image quality. When adaptive optics was first used in microscopy, it was common to rely on iterative approaches to determine the aberrations present. It is advantageous to avoid iteration, and therefore there has been a shift to deep learning for aberration prediction. However, issues remain regarding the practicalities of machine learning for adaptive optics, an important one being the requirement for a large training dataset. Here, we explore transfer learning to overcome this need for data by pre-training a network on a large simulated dataset and fine-tuning it with reduced experimental data for application in an experimental setting. We demonstrate that the pre-trained network can make noticeable improvements with fine-tuning on just 24 experimental samples. To further enhance practicality, we significantly extend the range of aberrations present, predicting up to 25 Zernike modes with each coefficient ranging from -1 to 1, and perform a thorough analysis of the type and magnitude of phase-diversity required in the input data for a successful network. Our approach demonstrates substantial aberration reduction on experimental data for 10 Zernike modes, with an average 73% decrease in RMS wavefront error from 1.81 to 0.48 rad when correction is applied. This method achieves complete experimental image capture and aberration inference at rates comparable to the image acquisition time of a typical laser scanning microscope. Additionally, we consider the benefits of further improvements via an iterative step. As such, this work addresses some of the key practical hurdles that remain in the use of deep learning for aberration prediction and correction.