Real-Time Sperm Detection Using Lightweight YOLOv5

Malformed sperm is an important cause of male infertility, and sperm morphology analysis (SMA) is an effective means to diagnose sperm morphology. Deep learning assists to enhance performance on precise SMA; however, existing deep learning based SMA methods mostly focus on single cell scale, which presents a challenge for obtaining single-sperm-image datasets (one sperm cell per image). It is also challenging to integrate current object detection models on low-performance devices. This paper presents a lightweight model for sperm detection. By removing 50% of convolutional kernels cutting the large-object-detecting head from YOLOv5s, our model got a similar precision to the original YOLOv3 (mAP.5 of 0.957 and 0.947, respectively), but with a model size of only 2.8 MB (123.6 MB of YOLOv3). There is a slight loss in precision compared to YOLOv5s (mAP.5:.95 of 0.604 with 14.4MB model size); however, our model still shows a significant advantage in reducing the number of parameters. Experimental results also indicated that MS COCO pre-training is helpful in sperm detection tasks, and the mosaic augmentation strongly enhances the precision for all YOLO models.