Enhancing single-molecule localization microscopy with deep learning

Single-molecule localization microscopy (SMLM) has significantly improved the visualization of sub-cellular structures, but enhancing the accuracy of 3D emitter localization remains challenging. The technique relies on precisely computationally localizing sparsely activated fluorophores, with traditional methods being iterative, time-consuming, and sensitive to camera noise and overlapping point spread functions (PSFs). We introduce a deep convolutional neural network that employs an innovative architecture to effectively manage diverse emitter scenarios, from isolated to densely packed. By transforming features from the real to the complex domain to integrate axial and lateral spatial information, our method outperforms existing deep learning-based localization algorithms. Tested on simulated SMLM frames with densities up to 2.0 µm⁻², our approach demonstrates superior performance across varying emitter densities and signal-to-noise ratios, maintaining high accuracy even under challenging conditions.