This paper presents a deep-learning method for deconvolution in light-sheet microscopy using propagation-invariant beams. Instead of end-to-end training with ground truths, the method leverages known physics of the imaging system. A generative adversarial network (GAN) is trained using images generated with the known point-spread function (PSF) and unpaired experimental data. The method achieves a two-fold improvement in image contrast compared to conventional methods, requiring only a few hundred regions of interest for training. Its effectiveness is demonstrated on various samples, including oocytes, preimplantation embryos, and brain tissue, using both Airy and Bessel beams.

Philip Wijesinghe, Stella Corsetti, Darren J. X. Chow, Shuzo Sakata, Kylie R. Dunning, Kishan Dholakia