摘要
目的探讨人工智能(AI)技术对视网膜静脉阻塞(RVO)的诊断价值。方法选取2016年7月至2022年6月在我院眼科就诊并进行免散瞳眼底照相的9000例患者图像,从中筛选出700例患有RVO的眼底图像,联合河南理工大学电气化学院,应用AI技术对眼底图像进行处理。本研究采用了最先进的两阶段算法(Faster-RCNN模型)、先进的一阶段检测算法(YOLOV4和YOLOV5模型),以及专门为RVO设计的改进型YOLOV5模型来处理收集到的RVO数据,并对结果进行对比分析。采用受试者工作特征曲线分析改进型YOLOV5模型对RVO诊断的灵敏度和特异度。结果Faster-RCNN模型检测精度最高,但其参数规模远大于其他模型,且每秒帧数仅为8,不能满足检测的实时性要求。改进型YOLOV5模型检测精度仅比Faster-RCNN模型低3%,但参数规模远远小于Faster-RCNN模型,且每秒帧数比其高22。根据不同的AI模型Faster-RCNN、YOLOV4、YOLOV5和改进型YOLOV5模型,绘制的受试者工作特征曲线下面积分别为0.946(95%CI:0.933-0.959),0.794(95%CI:0.772-0.816),0.864(95%CI:0.845-0.884),0.930(95%CI:0.915-0.944)。改进型YOLOV5模型对RVO诊断的灵敏度为87.0%,特异度为98.9%。结论改进型YOLOV5模型拥有较高的检测精确性,所需的参数规模较小,对RVO诊断的灵敏度和特异度高,可以作为RVO新的人工智能辅助诊断方法。
Objective To explore the diagnostic value of artificial intelligence(AI)in retinal vein occlusion(RVO).Methods The fundus images from 9,000 patients who visited our ophthalmology department and underwent non-mydriatic fundus photography from July 2016 to June 2022 were selected.Among these images,700 fundus images of patients with RVO were screened.These fundus images were processed using AI in collaboration with School of Electrical Engineering and Automation of Henan Polytechnic University.Besides,the state-of-the-art two-stage algorithms(the Faster-RCNN model),advanced one-stage detection algorithms(the YOLOV4 and YOLOV5 models),and a modified YOLOV5 model specifically designed for RVO were employed in this study to process the collected RVO data.In addition,a comparative analysis of the results was conducted.Moreover,the sensitivity and specificity of the YOLOV5 model for the diagnosis of RVO were evaluated based on the receiver operating characteristic(ROC)curve.Results The Faster-RCNN model demonstrated the highest detection accuracy;however,its parameter scale was significantly larger than that of other models,and its frame rate was only 8 frames per second,which did not meet real-time detection requirements.The detection accuracy of the modified YOLOV5 model was only 3%lower than that of the Faster-RCNN model,but it had a significantly smaller parameter scale compared with the Faster-RCNN model,achieving a frame rate of 22 frames per second higher than the Faster-RCNN model.The area under the ROC curve(AUC)plotted based on the Faster-RCNN,YOLOV4,YOLOV5,and modified YOLOV5 models was 0.946(95%CI:0.933-0.959),0.794(95%CI:0.772-0.816),0.864(95%CI:0.845-0.884),and 0.930(95%CI:0.915-0.944),respectively.The sensitivity and specificity of the modified YOLOV5 model for the diagnosis of RVO were 87.0%and 98.9%,respectively.Conclusion The modified YOLOV5 model,with its high detection accuracy,smaller parameter scale,and high sensitivity and specificity for the diagnosis of RVO,can be considered a new AI-assisted diagnostic method for RVO.
作者
刘松涛
吕辉
刘向玲
LIU Songtao;L Hui;LIU Xiangling(Department of Ophthalmology,Central Hospital of Jiaozuo Coal Group Co.,Jiaozuo 454000,Henan Province,China;School of Electrical Engineering,Henan Polytechnic University,Jiaozuo 454000,Henan Province,China;Department Of Ophthalmology,the Third Affiliated Hospital of Xinxiang Medical University,Xinxiang 453000,Henan Province,China)
出处
《眼科新进展》
CAS
北大核心
2025年第1期46-49,共4页
Recent Advances in Ophthalmology
基金
河南省科技攻关项目(编号:232102210171)
光电传感与智能测控河南省工程实验室资助项目(编号:HELPSIMC-2020-007)。
关键词
人工智能
视网膜静脉阻塞
医学图像分析
诊断价值分析
artificial intelligence
retinal vein occlusion
medical image analysis
diagnostic value analysis
