Abstract
The double-helix point spread function (DH-PSF) is a crucial tool in three-dimensional single-molecule localization microscopy (3D-SMLM). DH-PSF-based data analysis typically follows a two-step "detection-localization" workflow, where the detection stage predominantly relies on template matching (TM) methods. However, TM is highly sensitive to geometric deformations of the signal, a sensitivity that inherently conflicts with the physical characteristic of the DH-PSF rotating continuously along the axial depth, thereby constraining detection capability and processing speed. To overcome these bottlenecks, this paper proposes a novel localization algorithm, termed DH-DeepLoc, which deeply integrates a YOLOv5-based object detection network with a maximum likelihood estimation (MLE) multi-emitter fitting framework. Quantitative evaluations demonstrate that under various conditions of molecular density and signal-to-background ratio, DH-DeepLoc significantly outperforms traditional TM methods in terms of detection precision, recall, and processing speed, achieving an average detection speed approximately 25.53 times that of TM methods. Furthermore, comprehensive validation through both simulated data and real single-particle tracking experiments verifies that, compared to the classical EasyDHPSF algorithm, DH-DeepLoc not only maintains high-precision 3D localization performance but also exhibits superior robustness when handling dense and overlapping DH-PSF signals. By synergizing the ultra-fast detection advantages of deep learning with the high precision of MLE fitting, DH-DeepLoc provides an efficient, robust, and novel technical solution for 3D single-molecule localization imaging and multi-particle dynamic tracking.