Model-free machine learning-based 3D single molecule localisation microscopy

Single molecule localisation microscopy (SMLM) can provide two-dimensional super-resolved image data from conventional fluorescence microscopes, while three dimensional (3D) SMLM usually involves a modification of the microscope, for example, to engineer a predictable axial variation in the point spread function. Here we demonstrate a 3D SMLM approach (we call 'easyZloc') utilising a lightweight Convolutional Neural Network that is generally applicable, including with 'standard' (unmodified) fluorescence microscopes, and which we consider may be practically useful in a high throughput SMLM workflow. We demonstrate the reconstruction of nuclear pore complexes with comparable performance to previously reported methods but with a significant reduction in computational power and execution time. 3D reconstructions of the nuclear envelope and an actin sample over a larger axial range are also shown.