Super-resolution optical microscopy via a wavelet-spatial progressive network with high parameter efficiency

Super-resolution microscopy is indispensable for biomedical research, where deep learning has shown great promise. However, prevailing deep learning-based methods often compromise detail for artifact suppression, lack generalizability across microscopy modalities, and require massive parameters. To address these limitations, we present the wavelet-spatial progressive network (WSPN), which extracts details in the wavelet domain and suppresses artifacts in the spatial domain. Furthermore, we provide a publicly available confocal microscopy image dataset, BPAEC, which serves as an underexplored benchmark for evaluating cross-modality generalizability. We demonstrate that the WSPN has only 1 million parameters, reduced by 92% compared with the classical model, while maintaining a performance degradation of less than 3% on the established benchmark, BioSR, and showing effective cross-modality generalizability on the BPAEC. This combination of high-fidelity image quality and parameter efficiency makes WSPN a practical solution for super-resolution microscopy.