摘要目的:探讨基于眼前节光学相干断层扫描(AS-OCT)图像的核性白内障智能辅助分级算法对白内障分级的诊断价值。方法:采用诊断试验研究方法，收集2020年11月至2021年9月间同济大学附属第十人民医院电子病例系统中核性白内障患者939例1 608眼的AS-OCT图像资料，所有资料均符合临床阅片清晰度要求。其中男398例664眼，女541例944眼，年龄18～94岁，平均年龄(65.7±18.6)岁。由3名经验丰富的临床医生基于晶状体混浊分类系统(LOCS Ⅲ分级系统)，对所收集的AS-OCT图像进行1～6级人工标注。构建一种基于多级排序的全局-局部白内障分级算法，该算法包含5个基本的二元分类全局-局部网络(GL-Net)，每个GL-Net聚合白内障核区域、原始图像等多尺度信息进行核性白内障分级。基于消融实验和模型对比试验，采用准确率、精确率、灵敏度、F1指标及Kappa系数对模型性能进行评价，且所有结果均通过五折交叉验证。结果:模型在核性白内障数据集上的准确率、精确率、灵敏度、F1、Kappa分别为87.81%、88.88%、88.33%、88.51%、85.22%。消融实验结果表明，ResNet18结合局部特征和全局特征进行多级排序分类，模型在准确率、精确率、灵敏度、F1、Kappa指标上均有提升。与ResNet34、VGG16、Ranking-CNN、MRF-Net模型比较，本研究模型各性能指标均有提升。结论:基于深度学习的AS-OCT核性白内障图像多级排序分类算法对白内障分级具有较高的准确度，有望辅助提高眼科医生对核性白内障的诊断精度以及效率。

abstractsObjective:To investigate the diagnostic value of an intelligent assisted grading algorithm for nuclear cataract using anterior segment optical coherence tomography (AS-OCT) images.Methods:A diagnostic test study was conducted.AS-OCT image data were collected from 939 cases of 1 608 eyes of nuclear cataract patients at the Shanghai Tenth People's Hospital of Tongji University from November 2020 to September 2021.The data were obtained from the electronic case system and met the requirements for clinical reading clarity.Among them, there were 398 cases of 664 male eyes and 541 cases of 944 female eyes.The ages of the patients ranged from 18 to 94 years, with a mean age of (65.7±18.6) years.The AS-OCT images were labelled manually from one to six levels according to the Lens Opacities Classification System Ⅲ (LOCS Ⅲ grading system) by three experienced clinicians.This study proposed a global-local cataract grading algorithm based on multi-level ranking, which contains five basic binary classification global local network (GL-Net).Each GL-Net aggregates multi-scale information, including the cataract nucleus region and original image, for nuclear cataract grading.Based on ablation test and model comparison test, the model's performance was evaluated using accuracy, precision, sensitivity, F1 and Kappa, and all results were cross-validated by five-fold.This study adhered to the Declaration of Helsinjki and was approrved by Shanghai Tenth People's Hospital of Tongji University (No.21K216).Results:The model achieved the results with an accuracy of 87.81%, precision of 88.88%, sensitivity of 88.33%, F1 of 88.51%, and Kappa of 85.22% on the cataract dataset.The ablation experiments demonstrated that ResNet18 combining local and global features for multi-level ranking classification improved the accuracy, recall, specificity, F1, and Kappa metrics.Compared with ResNet34, VGG16, Ranking-CNN, MRF-Net models, the performance index of this model were improved.Conclusions:The deep learning-based AS-OCT nuclear cataract image multi-level ranking classification algorithm demonstrates high accuracy in grading cataracts.This algorithm may help ophthalmologists in improving the diagnostic accuracy and efficiency of nuclear cataract.