Automated sperm tracking using deep learning for high-resolution analysis of motility

Abstract

Sperm motility is a key determinant of cell function, and it is critical for male fertility. In the analysis of human semen, motility is routinely assessed using manual evaluation or computer-assisted sperm analysis (CASA). However, these approaches provide limited access to single-cell dynamics, use proprietary software, and rely on population- and time-averaged descriptors of sperm motion that can obscure both the intrinsic heterogeneity and the dynamic behaviors within individual trajectories. Here, we present an open-source, modular framework for automated sperm detection and tracking that enables trajectory-resolved quantification across species. The pipeline combines deep learning–based sperm head segmentation using species-specific Cellpose models with automated tracking via Trackpy, enabling reconstruction of individual sperm trajectories from time-lapse microscopy. This workflow was applied across four species: human, mouse, rat, and bovine. In human sperm, it was further tested under both basal and hyperactivation-promoting conditions. Detection performance was consistently high across species, with F1 scores, a metric that measures model accuracy, between 0.96 and 0.98 on independent validation datasets. From the obtained trajectories, CASA-like parameters, such as curvilinear velocity, straight-line velocity, and amplitude of lateral head displacement, were computed, showing strong agreement in human sperm with conventional CASA. An automated Python-based software for sperm tracking was designed to facilitate access to the technology. Overall, this framework provides a transparent and flexible alternative to CASA, enabling high-resolution, single-cell analysis of heterogeneous sperm motility and offering new opportunities to study sperm function.