A cascaded deep network for reconstruction of structured illumination microscopy

Abstract

Structured illumination microscopy (SIM) is one of the significant super-resolution (SR) optical microscopic techniques most suitable for live-cell imaging. Here, we devise a cascaded deep network (entitled CAS-Net) for super-resolution reconstruction of SIM. The CAS-Net comprises two subnets, the first uses a mixed Fourier attention transformer network (MFATN) to reconstruct a SR image from nine raw images, while the second incorporates sparsity constraint with a U-Net and further improves the resolution to ∼80 nm. Particularly, the subnet that provides physical and sparsity constraints circumvents the need for manually labeled datasets and the necessity to tune manually the fitting parameters required in the conventional sparse deconvolution process. The generality of the proposed CAS-Net was tested with the raw images acquired by different SIM microscopes. We can envisage that the CAS-Net will be widely applied for SIM techniques in the future.