ヒト精子濃度、運動性、運動学的分析のための人工知能ベースのコンピュータビジョンモデル

Accurate assessment of sperm concentration and motility is critical for the diagnosis and management of male infertility. However, current methods, manual hemocytometer counting and commercial computer-aided sperm analysis (CASA) systems, are limited by labor intensity, human error, and variable performance under diverse sample conditions. Here, we present an artificial intelligence (AI)-driven computer vision tool for high-resolution, quantitative analysis of sperm motility and concentration. In a prospective study of 26 semen samples (22 patients, 4 donors), we benchmarked the AI model against manual tracking (using Fiji software) and a commercial CASA system (Hamilton Thorne IVOS II). Our method computed concentration and motility parameters, including straight-line velocity (VSL), curvilinear velocity (VCL), average path velocity (VAP), linearity (LIN), amplitude of lateral head displacement (ALH_max), and beat cross frequency (BCF). Calibration using donor samples enabled accurate mapping of tracked sperm counts to concentrations. The AI tool presented a strong linear correlation with manual tracking (R ² = 0.93-0.98; Root Mean Square Error (RMSE) = 3.3-7.3 μm/s for VSL, VCL, VAP), and outperformed CASA in both accuracy and consistency across all motility parameters. Post-calibration, ALH_max and BCF estimates improved substantially, with a 30%-50% reduction in RMSE. Grading of sperm motility by the AI model aligned closely with manual classification, avoiding the systematic misclassification typically observed with CASA. Furthermore, the AI system exhibited higher repeatability and robustness across duplicate samples and variable imaging conditions, with deviations below ± 2%. These findings demonstrate that our AI-based tool offers a quantitative and reliable alternative to current semen analysis platforms, supporting improved fertility diagnostics and potentially a more informative treatment process.