Volumetric single-molecule localization imaging with system-aware deep learning

Super-resolution fluorescence imaging has promoted the investigation of cellular morphology and mechanisms that are concealed by the diffraction limit. Among all techniques, localization microscopy has emerged as one of the most adapted approaches due to its supreme resolution performance. However, a careful balance between hardware complexity, computational load, and user accessibility is strongly needed. Traditional approaches often rely on either deterministic models or deep learning algorithms, missing the potential benefits of combining both. In this study, we present deep volumetric localization microscopy (VLM), a novel super-resolution imaging approach that merges optical advancement with computational innovations to achieve high-fidelity 3D single-molecule localization. VLM employs a wavefront-optimized light-field setup to efficiently record single-molecule emissions, which are then processed by cascaded neural networks to reconstruct 3D volumes and localize molecular positions with precisions of 10 nm laterally and 25 nm axially, over imaging depths exceeding 4 μm. VLM adopts a unique system-aware training strategy by synthesizing the training data with intrinsic point-spread functions, thereby eliminating the need for additional imaging references and sample dependencies. This design enhances robustness and generalizability across different biological contexts. We demonstrate VLM’s performance across biological specimens, showcasing its super-resolution performance with minimal phototoxicity, system simplicity, and efficient data acquisition. We believe that VLM will address the challenges in superresolution microscopy, expand the applicability of single-molecule imaging, and facilitate discoveries in cell biology, disease mechanisms, and biomedical innovation.