Deep Characterization For Sperm Analysis: Leveraging A Novel Dataset To Enhance Motility Assessments

Microscopic imaging techniques have been developed to assess sperm quality by analysing morphology, motility and other parameters. The study presented in this paper introduces an automatic computer-assisted analysis (CASA) system driven by deep learning. It covers species classification, sperm detection and tracking, and sample quality analysis. Our first contribution is the introduction of a new curated and well-labeled dataset specifically designed for sperm analysis, addressing the crucial need for diverse data in this field. To the best of to our knowledge, there exists no extensive dataset encompassing a wide variety and volume of sperm species. The proposed dataset will thus serve as a substantial contribution in this domain. The second contribution consists in the training and testing of state-of-the art deep-learning models (MobileNet, YOLOv5s, and DeepSort) on our proposed data sets as well as on of the widely used VISEM dataset [1]. Extensive assessment for classification, detection, and tracking has yielded promising results: classification accuracy of 99.2%, detection results ranging from 96% to 99.5% mAP@50, and a tracking MOTP of 99%. Rigorous testing across various sperm species provides vital parameters for sperm motility assessment.