3D Particle Localization in Living Cells by Deep Learning

Point spread function engineering techniques have extended the axial range accessible to fluorescence microscopy by simple modifications to the optical path of a microscope. A particularly promising engineered PSF is the double helix PSF (DHPSF), which provides an axial range of over 2 μm. This extended depth allows sub-diffraction localization of point sources in 3D across an axial range approaching the typical thickness of mammalian adherent cells. Thus, DHPSF is not only attractive for PALM-style superresolution experiments but also for tracking small punctate structures in cells. Still, data analysis of DHPSF images presents a challenge because no analytic form of the DHPSF exists. This difficulty has been addressed previously by using heuristic functions for fitting or interpolations of calibration scans of fluorescent beads. However, this approach is limiting for cellular particle tracking applications of DHPSF because of the heterogeneous signal background, high particle densities, and large variation in intensities of the target particles encountered in such experiments. Here, we present a data-driven and fast method based on deep learning (DL) to address these challenges. Specifically, we use the DL approach to analyze DHPSF data without the need to introduce heuristic fitting functions or interpolate calibration scans. We demonstrate the accuracy of the method on localizing simulated data in 3D, and test its efficacy in the presence of noisy, heterogeneous backgrounds. Finally, we localize and track in 3D puncta of human T-cell leukemia virus type 1 (HTLV-1) in live cells and present preliminary data on the mobilities of these particles measured in this noisy environment. This work has been supported by grants from the National Institutes of Health (R01 GM098550 and RO1 GM124279).