Fluorescence microscopy enables high-contrast visualization of subcellular structures through the use of synthetic fluorescent markers, though it is limited by lower spatial resolution and the risk of photobleaching. Super-resolution (SR) fluorescence microscopy overcomes the diffraction barrier to resolve finer details, but comes with trade-offs such as decreased temporal resolution, the need for sophisticated equipment, optical sectioning requirements, and constraints on the sample’s exposure tolerance.
In this study, we introduce a deep learning model that employs upsampling and downsampling to achieve super-resolution. By connecting up-projection and down-projection blocks in a cascaded arrangement, our model maintains the important relationships between the input low-resolution (LR) images and the target high-resolution (HR) labels. To extract dependencies across various depths, feature maps produced at intermediate upsampling stages are combined. Additionally, external self-attention blocks are integrated after these intermediate up-projection blocks to assess pixel-level affinities. Unlike traditional self-attention mechanisms that focus on correlations within a single sample, our approach uses external attention to capture inter-sample correlations across the entire dataset, offering the added benefit of reduced computational demands.