Sperm Detection Based on You Only Look Once-Neural Architecture Search (YOLO-NAS)

Abstract

Computer-Aided Sperm Analysis (CASA) systems are vital for male reproductive diagnostics. However, the CASA systems are prone with detection challenges. Therefore, this study explores You Only Look Once-Neural Architecture Search (YOLO-NAS) for precise, rapid sperm analysis, addressing its underexplored efficacy in this domain. Using the Sperm Videos and Images Analysis (SVIA) dataset (3,590 images, 125,000 objects), XML annotations were converted to YOLO TXT format. YOLO-NAS models (S, M, L) were trained for 40 epochs on Google Colab Pro using AdamW, batch size 32, and learning rates 0.002 (initial 0.0005). The performance was evaluated using mAP@0.50, Recall@0.50, Precision@0.50, and F1 Score@0.50, with an IoU threshold of 50%. The YOLO-NAS-L was most effective for sperm/impurity detection, achieving 0.8163 mAP@0.50 and 0.8203 precision@0.50. Increased model complexity enhances feature extraction, especially for small objects while higher precision may decrease recall. From the outputs, this study has successfully developed and evaluated a YOLO-NAS model for sperm detection. Future works include more epochs, diverse datasets, and offline training in sperm clustering, as well morphology and motility analyses.