深度学习在后发性白内障混浊分析中的应用研究  

The application of deep learning in the analysis of posterior capsule opacity after cataract

作　　者：胡师尧 陈媛媛 李辰 严宏 Shiyao Hu;Yuanyuan Chen;Chen Li;Hong Yan(Doctoral degree 2020,Faculty of Electronic and Information Engineering,Xi'an Jiaotong University,Xi'an 710049,China;Shaanxi Eye Hospital,Xi'an People's Hospital(Xi'an Fourth Hospital),affiliated People's Hospital of Northwest University,Xi'an 710004;Faculty of Electronic and Information Engineering,Xi'an Jiaotong University,Xi'an 710049,China)

机构地区：[1]西安交通大学电信学部计算机科学与技术学院,710049 [2]西安市人民医院(西安市第四医院)眼科陕西省眼科医院,710004

出　　处：《中华眼科医学杂志(电子版)》2024年第5期262-268,共7页Chinese Journal of Ophthalmologic Medicine(Electronic Edition)

基　　金：国家自然科学基金(81873674);陕西省重点研发计划项目(2021ZDLSF02-08);西安市创新能力强基计划(21YXYJ0005)。

摘　　要：目的探索深度学习在人工晶状体(IOL)后囊混浊分析中的应用。方法收集2020年9月至2023年7月于陕西省眼科医院就诊行白内障摘除联合IOL植入术62例(100只眼)患者的术后数月至数年采用裂隙灯显微镜后照法拍照图像100张。其中,男性24例(34只眼),女性38例(66只眼);年龄41~78岁,平均年龄(58.7±9.82)岁。晶状体后囊混浊分析框架包含IOL区域分割、IOL中心定位、混浊区域分割及混浊表征提取等四个主要模块。其中,IOL区域分割和混浊区域分割均采用Unet作为分割模型,所使用的数据集包含100张白内障术后IOL眼后囊膜的裂隙灯后照法图片;IOL中心定位使用几何矩算法计算;混浊表征提取使用主干为Resnet模型对患者的3个视觉质量分析系统视力指标进行预测。以测试集人工标注标签为参照,评估模型IOL区域分割的结果,使用Python软件计算交并比(IoU)、Dice系数及召回率。以测试集人工标注标签为参照,评估模型混浊区域分割的结果,使用Python软件计算准确率、精确率、召回率及f1-score。以测试集视觉质量分析系统(OQAS)指标真实值为参照,评估模型的图像回归预测结果,使用Python软件计算平均绝对误差。结果20张测试集IOL区域分割的IoU为0.9117,Dice系数为0.9527,召回率为0.9524,f1-score为0.9527。20张测试集混浊区域分割的平均准确率为0.9690,平均精确率为0.9329,平均召回率为0.9264,平均f1-score为0.9191。训练完成的模型在7张测试集上进行消融,原图+热力图+混浊区域、原图+热力图、原图+混浊区域及仅原图指导后,模型预测调制传递函数、斯特列尔比、客观散射指数的平均绝对误差分别为2.4319、0.0154、3.4032、4.3300、0.0166、2.9997、10.5013、0.0161、2.8775、3.8151、0.0195及3.7067。在输入中单独加入热力图和混浊区域时,模型预测客观散射指数的平均绝对误差小于其他模型变体。结论深度学习模型在晶状�Objective The aim of this study is to explore the application of deep learning in the analysis of posterior capsule opacification(PCO)following intraocular lens(IOL)implantation.Methods Slit-lamp retroillumination images of 100 eyes from 62 patients who underwent cataract extraction combined with IOL implantation at the Shaanxi Eye Hospital between September 2020 and July 2023 were collected.The cohort included 24 males(34 eyes)and 38 females(66 eyes),with an average age of(58.7±9.82)years(ranging from 41 to 78 years).The PCO analysis framework consisted of four main modules:IOL region segmentation,IOL center localization,opacification region segmentation,and extraction of opacification features.Both the IOL region segmentation and opacification region segmentation modules employed the U-Net model,trained on a dataset of 100 slit-lamp retroillumination images of posterior capsules with implanted IOL.The IOL center was localized using the geometric moment algorithm,while opacification feature extraction used a ResNet-based model to predict three visual quality metrics of patients.The performance of the IOL region segmentation model was evaluated against manually labeled test set annotations.Metrics such as Intersection-over-Union(IoU),Dice coefficient,and recall rate were calculated using Python software.The opacification region segmentation results were similarly evaluated with metrics including accuracy,precision,recall,and f1-score.For image regression tasks predicting OQAS(Optical Quality Analysis System)metrics,the mean absolute error(MAE)was computed using Python software.Results In the IOL region segmentation task,the test set achieved an IoU of 0.9117,Dice of 0.9527,recall of 0.9524,and f1-score of 0.9527.In the opacification region segmentation task,the test set achieved an average accuracy of 0.9690,precision of 0.9329,recall of 0.9264,and f1-score of 0.9191.In the visual indicator prediction task,the mean error of strehl ratio,modulation transfer function,object scatter index were2.4319,0.0154,3.4032.C

关 键 词：晶状体后囊混浊 深度学习 裂隙灯图像 医学图像处理 

分 类 号：R776.1[医药卫生—眼科]