摘要
目的:针对已有算法无法在超声生物显微镜(ultrasonic biological microscopy,UBM)图像中同时测量多个眼前节生理参数的问题,提出一种基于YOLOv5的眼前节参数实时测量算法。方法:先使用YOLOv5算法在UBM图像中定位出中央角膜上皮层、中央角膜内皮、两侧前房角、两侧睫状沟、瞳孔两端、瞳孔中央及晶状体后囊,再结合先验知识将其排序并筛选出合格图像后,计算中央角膜厚度、前房深度、晶状体厚度、睫状沟间距、前房角间距及瞳孔直径等眼前节生理参数。使用欧氏距离之差的绝对值评估该算法对眼前节生理参数的测量误差。结果:该算法对眼前节关键点的定位误差为(66.27±66.25)μm,预测中央角膜厚度的相对误差为9.61%,预测其他眼前节参数的相对误差均在3%以下。在配备Ryzen 5 4600U处理器的便携式计算机上检测一张UBM图片仅需要140 ms。结论:该算法准确性高、实时性强、对硬件要求低,能够精确且快速地测量多个眼前节生理参数。
Objective To propose a YOLOv5-based real-time measurement algorithm for anterior segment parameters to eliminate the deficiencies of the existing algorithms for detecting multi anterior segment parameters in ultrasonic biological microscopy(UBM)images.Methods The central corneal epithelium,central corneal endothelium,anterior chamber angles on both sides,ciliary sulci on both sides,both ends of the pupil,the center of the pupil and the posterior capsule of the lens in the UBM images were located with YOLOv5 algorithm and then sorted with priori knowledge to pick out the qualified images,and the anterior segment parameters were calculated including the central corneal thickness,anterior chamber depth,lens thickness,ciliary sulcus spacing,anterior chamber angle spacing and pupil diameter.Results The algorithm proposed had the positioning error of the anterior segment being(66.27±66.25)μm,the relative error of the central cornea thickness being 9.61%and the prediction relative error of other anterior segment parameters below 3%,which took only 140 ms to measure a UBM image on a portable computer equipped with a Ryzen 54600U processor.The measurement error of the algorithm for the anterior segment parameters was evaluated using the absolute value of the difference between Euclidean distance.Conclusion The algorithm proposed gains advantages in high accuracy,timeliness and low requirements for hardware,and can be used to detect multi anterior segment parameters rapidly.
作者
贺永强
齐珊
李泽萌
周盛
杨军
林松
HE Yong-qiang;QI Shan;LI Ze-meng;ZHOU Sheng;YANG Jun;LIN Song(Institute of Biomedical Engineering,Chinese Academy of Medical Sciences,Peking Union Medical College,Tianjin 300192,China;Tianjin Maida Medical Technology Co.,Ltd.,Tianjin 300384,China;Eye Hospital of Tianjin Medical University,Tianjin 300384,China)
出处
《医疗卫生装备》
CAS
2023年第6期7-13,共7页
Chinese Medical Equipment Journal
关键词
YOLOv5
眼前节参数
超声生物显微镜
目标检测
深度学习
YOLOv5
anterior segment parameter
ultrasonic biological microscopy
object detection
deep learning
