A parametric synthetic data generator for training learning-based sperm analysis systems

Computer-Aided Sperm Analysis (CASA) systems are used as a primary tool to evaluate male fertility. While they are widely employed for treating human infertility, they are also well-known in domestic animal reproduction, where they are applied to conduct sperm competition studies and wildlife semen assessment. The lack of publicly available data sets presents a major challenge to train the models that are used by this systems. In this paper, we propose a synthetic data generator based on a parametric model. To validate our approach, we evaluate two key tasks in sperm analysis: sperm count and sperm motility assessment. Our proposed system is trained with synthetic data and is validated on real-world boar sperm images. The results obtained by our system show a mean average precision (mAP) of 84.3% for sperm counting and a multiple object tracking accuracy (MOTA) of 80.5% for motility estimation. Furthermore, our solution requires minimal computational resources, allowing its use on embedded devices, which are easier to integrate into professional CASA systems.