Improving Reconstruction of Structured Illumination Microscopy Images via Dual-Domain Learning

Abstract

Structured illumination microscopy (SIM) provides an enhanced resolving power surpassing the optical diffraction limit by optical modulation of patterned illuminations. Although end-to-end deep learning techniques have recently advanced the reconstruction of SIM images, the reconstruction fidelity of existing networks is still moderate. We experimentally point out the crux lies in the inability of these models for faithful frequency learning. As a remedy, we propose a dual-domain learning strategy for SIM reconstruction, namely DDL-SIM, which learns to reconstruct SIM images from raw images in the spatial domain and raw image spectra in the frequency domain simultaneously, with the goal of narrowing the reconstruction gaps in both domains, thereby better recovering modulated frequencies and resolving more fine structures. Reconstruction experiments across various biological structures demonstrate the proposed DDL-SIM significantly improves the reconstruction fidelity of SIM images and shows great robustness against reconstruction artifacts.