A framework for evaluating predicted sperm trajectories in crowded microscopy videos

Since the 1980s, semi-automated sperm motility analysis of phase contrast microscopy videos has been used to measure and categorize sperm motility patterns. Motility categories are determined from various kinematic parameters such as Curvilinear Velocity (VCL) and Beat Cross Frequency (BCF). These measures ultimately rely on the quality of the tracking for each individual sperm in the microscopy video. However, common approaches to sperm tracking require sample dilution and shortening the time window of observation (less than 1 to 2 seconds) to avoid tracking errors that occur when sperm cross paths. The post-ejaculatory lifespan of sperm can exceed several hours to days in some species, and long-term adaptive changes in motility pattern may be an important distinguishing factor for predictive modeling of sperm fertilizing competence. Improving the predictive value of computer assisted semen analysis will require accurate tracking of sperm trajectories over physiologically-relevant time scales and at the high cell densities typically found in semen. In this work, we identify a framework for accurately assessing the quality of sperm trajectory tracking that is independent of standard motility measures. We utilize cell tracking metrics adapted from the more common task of tracking adherent somatic cells and propose modifications based on the unique challenges of sperm video-microscopy. We also provide a small dataset of microscopy videos that includes 340 labeled sperm trajectories to allow for future comparisons and developments. Finally, we demonstrate that variations in configuration can lead to as much as a 30% improvement on metrics, showcasing their effectiveness at analyzing tracking quality.Author summary: This report develops a computer vision framework to track individual sperm cells in crowded (high cell-density) microscopy videos, with the goal of improving automated analysis of sperm motility patterns. The task of accurately analyzing sperm movement is deceptively challenging due to the high rate of cell crossovers, an issue that has long impeded long-term tracking of sperm trajectories. Here, we introduce new evaluation metrics and provide a labeled dataset to serve as a baseline for future improvements.