Improved YOLOv8-based measurement method of anterior segment parameters of ultrasonic biological microscopy images
10.19745/j.1003-8868.2024022
- VernacularTitle:基于改进YOLOv8的生物显微镜图像眼前节参数测量方法研究
- Author:
Yong-Qiang HE
1
;
Qing-Hao MIAO
;
Cai-Lian XIONG
;
Jun YANG
Author Information
1. 中国医学科学院北京协和医学院生物医学工程研究所,天津 300192
- Keywords:
YOLOv8;
ultrasonic biological microscopy;
anterior segment parameter;
parameter measurement;
generative ad-versarial network
- From:
Chinese Medical Equipment Journal
2024;45(2):8-16
- CountryChina
- Language:Chinese
-
Abstract:
Objective To propose a measurement method based on improved YOLOv8 for the anterior segment parameters of ultrasonic biological microscopy(UBM)images to solve the problems of the ophthalmic ultrasound images in low quantity,high annotation cost and weak model generalization ability.Methods Firstly,data enhancement using StyleGAN3 was carried out to improve YOLOv8 model in decreased sensitivity to UBM images and increased generalization ability;secondly,the pseudo-labels of virtual UBM images were generated based on the YOLOv8 model with the original UBM images as the dataset and the method of pseudo-label self-training in semi-supervised learning;finally,the YOLOv8 model was trained with the expan-ded dataset and improved with the global context network(GCNet)module,and the prediction results of the YOLOv8 model were sorted using the key-point ranking algorithm,and the measurements of the anterior segment physiological parameters were computed after screening qualified images based on prior knowledge.Results Compared with the hand-labeled results by the ophthalmologists,the localization error of the StyleGAN3 data-enhanced and self-trained YOLOv8 model was(61.94±40.66)μm,and the mean relative errors for the measurements of anterior chamber angle distance,pupil diameter,ciliary sulcus distance,central corneal thickness,anterior chamber depth and lens thickness were 0.62％,1.35％,0.68％,4.87％,0.93％and 0.75％,respectively.Conclusion The method proposed enhances the accuracy of the measurement method for anterior segment parameters of UBM images,and can meet real-time requirements.[Chinese Medical Equipment Journal,2024,45(2):8-16]
