The quality of images created by diffraction-limited light microscopy modalities (e.g. most widefield, confocal and light sheet) is deteriorated by the point spread function (PSF) of the system. Using the PSF, classical deconvolution methods, such as the Richardson–Lucy algorithm, can be applied to computationally restore images. The PSF can be obtained by 1) direct measurement from the images or 2) estimation. Neither solution is ideal as it requires either a laborious manual measuring process or expert level knowledge of many hardware components that affect the modeling function.

With deep learning (DL), the PSF is directly learned from easy to generate training images - no need to measure it or deterministically model it. Training of a DL model is an iterative process which can be efficiently done by using multiple GPUs, e.g. via Aivia Cloud. Once created, it can be used on new images in a single step. By contrast, with classical deconvolution the restoration process is iterative for every image one desires to process.