The 3D point spread function can be autoencoded in latent space

Abstract

The 3D point spread function of a fluorescence emitter in a living cell is often different from that of the z-stack of bead images typically used as a reference. Here we show that the location in z of a fluorescent emitter can be directly accessed in the latent space of an autoencoder trained on the sample images without z-reference data. The corresponding decoding network represents an accurate point spread function.