摘要
针对白带显微图像中白细胞、念珠菌、滴虫的自动化识别问题,提出基于改进的YOLOv3-tiny与Resnet50相结合的细胞识别算法。在细胞检测阶段,基于改进的YOLOv3-tiny算法进行细胞检测,结合显微图像中细胞均为小目标的特点,提高对细胞的检测准确率。在细胞识别阶段,采用Resnet50模型对细胞进行精确分类,并结合显微图像细胞数据集的特点,运用迁移学习的方法进行模型训练,提高分类的准确率。实验结果表明,该算法对白细胞、念珠菌、滴虫的识别结果和效率均优于其他算法,在319例白带样本上的测试结果表明:白细胞的灵敏度、特异性、符合率分别为90.75%、93.84%、92.16%,念珠菌的灵敏度、特异性、符合率分别为95.05%、96.33%、95.92%,滴虫的灵敏度、特异性、符合率分别为94.74%、98.00%、97.81%,达到临床检验的要求。
Aiming at the automatic recognition of white blood cell,monilia and trichomonad of leucorrhea microscopic image,this paper proposes the cell detection algorithm based improved YOLOv3-tiny algorithm and Resnet50 network.In phase of cell detection,cells were detected based on the improved YOLOv3-tiny algorithm,and the characteristics that cells in the microscopic images were small objects were combined to improve the accuracy of detection.In phase of cell recognition,Resnet50 network was used for cell classification.Considering the features of microscopic images datasets,we use transfer learning for model training to improve the classification accuracy.The experimental results show that the recognition results and efficiency of white blood cell,monilia and trichomonad of our algorithm are better than other algorithms.The results of 319 leucorrhea samples show that the sensitivity,specificity and accuracy of white blood cell are 90.75%,93.84%and 92.16%respectively;the sensitivity,specificity and accuracy of monilia are 95.05%,96.33%,95.92%;the sensitivity,specificity and accuracy of trichomonad are 94.74%,98.00%,97.81%.The results can meet the requirement of clinical testing.
作者
侯剑平
王超
赵万里
段忆芮
Hou Jianping;Wang Chao;Zhao Wanli;Duan Yirui(Autobio Labtec Instruments Co.,Ltd.,Zhengzhou 450016,Henan,China)
出处
《计算机应用与软件》
北大核心
2021年第9期232-238,共7页
Computer Applications and Software
