Sperm Detection and Tracking Model Using HDE Transformer With Spatio-Temporal Deformable Attention for Sperm Analysis Automation

Infertility is a common problem, affecting approximately one in six adults worldwide. Several studies have shown that the male factor is responsible for infertility in up to 50% of couples. Elaborate sperm analysis is needed to solve this problem. Manual evaluation of sperm motility requires precise microscopic observation. In addition, the rapid movement of sperm in the microscopic field of view requires a large amount of human and time labor. To address this issue, the adoption of Computer-Assisted Sperm Analysis (CASA) systems has been increasing in clinical settings as a replacement for manual evaluations conducted by experts. To automate sperm analysis, it is necessary to detect and track target sperm from microscopic videos. However, a baseline has not been established for an end-to-end model that performs sperm detection and tracking simultaneously. To contribute to the automation of sperm evaluation, we propose HDE-Track, a model that considers both the detection of blurry sperm and the movement of sperm. Through experiments conducted using VISEM-Tracking(a human spermatozoa tracking dataset), HDE-Track achieved 2.0% improvement in HOTA and 5.2% improvement in MOTA compared to the existing model TransTrack. The proposed method, HDE-Track, contributes to the automation of sperm analysis by integrating with conventional CASA systems. HDE-Track can serve as a stronger baseline for future research on temporal modeling and Transformer-based sperm detection and tracking models.