Cellular fluorescence microscopy aberration correction with focus-diverse machine learning

A microscope objective is a significant cost of an optical system, but a low-cost objective may introduce optical aberrations into the final image, making specimen interpretation difficult. Our research investigates using focus-diverse image stacks and machine learning to correct aberrations computationally and restore perceptual similarity to the aberrated image. Our previous research showed promising results for entirely synthetic cell datasets, but lacked a way to viably extend the method to experimental data because of the training data required. This research focuses specifically on a method to artificially aberrate experimental data and generate a semi-synthetic dataset to support focus-diverse machine learning efforts. Performance is evaluated with quantitative and qualitative metrics including peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) index. Qualitative and quantitative results show that artificial aberration is a viable method for generating semi-synthetic datasets, and that focus-diverse machine learning trained on semisynthetic datasets can be used for aberration correction. This is an important step to able generate the quantity of focus-diverse aberrated training data required by machine learning models.