Physics-Inspired Single-Particle Tracking Accelerated with Parallelism

Abstract

Traditional particle tracking methods approximate each pixel as reporting a continuous grayscale intensity, an assumption valid for long exposures with many detected photons. At short exposures, however, pixels record only a few discrete photon counts, and this quantization introduces both statistical and computational challenges. We present QTrack, a likelihood-based and parallelized framework that directly models discrete photon detections while exploiting correlations across all frames. By doing so, QTrack surpasses the Cramer - Rao lower bound (CRLB) prediction assuming continuous intensity and localization-and-linking. To make this possible in practice, QTrack exploits the parallelism inherent in both likelihood evaluation and posterior sampling through vectorization, multi-threading, and GPU acceleration, with lightweight interthread communication. On a single mid-range GPU (Nvidia GTX 1060, 6 GB), QTrack achieves up to a 50-fold speedup compared to a serial CPU implementation, while maintaining full accuracy and data efficiency. Together, these advances establish that short exposures with quantized photon detections are not a limitation but an opportunity: when modeled rigorously, they enable localization and diffusion coefficient estimates beyond the CRLB prediction for continuous-intensity data, setting a new standard for single-particle tracking.