The team introduced a physics-informed masked autoencoder, PiMAE, to learn the optical imaging mechanism directly from raw microscope observations.

The self-supervised method treats the point-spread function as a hidden pattern and fluorescent emitters as a sparse representation. It recovered the point-spread function from as few as five raw images, without additional annotation.

Tests on synthetic data showed improved reconstruction accuracy, while experimental wide-field images approached the results of structured illumination microscopy. The work appeared in Photonics Research. Ze-Hao Wang and Long-Kun Shan are first authors; Fang-Wen Sun is the corresponding author.