Developing Noise-Resistant Three-Dimensional Single Particle Tracking Using Deep Neural Networks

Three-dimensional single particle tracking (3D SPT) is a powerful tool in various chemical and biological studies. In 3D SPT, z sensitive point spread functions (PSFs) are frequently used to generate different patterns, from which the axial position of the probe can be recovered in addition to its xy coordinates. Conventional linear classifier-based methods, for example, the correlation coefficient method, perform poorly when the signal-to-noise ratio (S/N) drops. In this work, we test deep neural networks (DNNs) in recognizing and differentiating very similar image patterns incurred in 3D SPT. The training of the deep neural networks is optimized, and a procedure is established for 3D localization. We show that for high S/N images, both DNNs and conventional correlation coefficient-based method perform well. However, when the S/N drops close to 1, conventional methods completely fail while DNNs show strong resistance to both artificial and experimental noises. This noise resistance allows us to achieve a camera integration time of 50 μs for 200 nm fluorescent particles without losing accuracy significantly. This study sheds new light on developing robust image data analysis methods and on improving the time resolution of 3D SPT.