目的 设计一套基于深度学习的自动化检测流程,评估该流程在形态检测的效率、准确性、可靠性.方法 使用预训练YOLO目标检测模型和VGG16分类模型分析1 000例样本的各14张图片,每例样本至少分析200条精子,人工镜检等量样本,对比两方法效率、准确性、相关性.结果 使用预训练的分类模型对人工分类但未训练的精子头进行形态检测,预测准确性达到95.5%,临床每例样本检测时间仅用10 s,准确性与效率高于人工镜检,两种检测方法的正常形态精子百分率呈显著正相关(r=0.84,P<0.01).结论 本研究设计的检测流程能极大程度地提高工作效率,可靠性、准确性超过人工镜检.
Objective To design an automated detection protocol for morphological detection of sperm head based on deep learning and evaluate its efficiency,accuracy and reliability.Methods Fourteen pictures for each of 1 000 samples were analyzed by using the pre-trained YOLO target detection model and VGG16 classification model.At least 200 sperm were detected for each sample.Equal amount of samples were analyzed by manual microscope examination,and the efficiency,accuracy and correlation between the two methods were compared.Results The morphology of sperm heads which were manually classified but untrained was detected by pre-trained classification model,and the prediction accuracy reached to 95.5%.The detection time for clinical each sample was only 10 seconds,and its accuracy and efficiency were higher than those of manual microscope examination.The percentages of the sperms with normal morphology were significantly positively correlated(r=0.84,P<0.01)in the detections of both the methods.Conclusion The detec-tion protocol proposed in this study can greatly improve work efficiency,and its reliability and accuracy exceed those of manual micro-scope examination.