Deep learning for precise axial localization of trapped microspheres in reflective optical systems

High-precision axial localization measurement is an important part of micro-nanometer optical measurement, but there have been issues such as low calibration efficiency, poor accuracy, and cumbersome measurement, especially in reflected light illumination systems, where the lack of clarity of imaging details leads to the low accuracy of commonly used methods. Herein, we develop a trained residual neural network coupled with a convenient data acquisition strategy to address this challenge. Our method improves the axial localization precision of microspheres in both reflective illumination systems and transmission illumination systems. Using this new localization method, the reference position of the trapped microsphere can be extracted from the identification results, namely the "positioning point" among the experimental groups. This point relies on the unique signal characteristics of each sample measurement, eliminates systematic repeatability errors when performing identification across samples, and improves the localization precision of different samples. This method has been verified on both transmission and reflected illumination optical tweezers platforms. We will bring greater convenience to measurements in solution environments and will provide higher-order guarantees for force spectroscopy measurements in scenarios such as microsphere-based super-resolution microscopy and the surface mechanical properties of adherent flexible materials and cells.