PURPOSE To develop a deep learning approach based on deep residual neural network (ResNet101) for the automated detection of glaucomatous optic neuropathy (GON) using color fundus images, understand the process by which the model makes predictions, and explore the effect of the integration of fundus images and the medical history data from patients. METHODS A total of 34,279 fundus images and the corresponding medical history data were retrospectively collected from cohorts of 2371 adult patients, and these images were labeled by 8 glaucoma experts, in which 26,585 fundus images (12,618 images with GON-confirmed eyes, 1114 images with GON-suspected eyes, and 12,853 NORMAL eye images) were included. We adopted 10-fold cross-validation strategy to train and optimize our model. This model was tested in an independent testing dataset consisting of 3481 images (1524 images from NORMAL eyes, 1442 images from GON-confirmed eyes, and 515 images from GON-suspected eyes) from 249 patients. Moreover, the performance of the best model was compared with results obtained by two experts. Accuracy, sensitivity, specificity, kappa value, and area under receiver operating characteristic (AUC) were calculated. Further, we performed qualitative evaluation of model predictions and occlusion testing. Finally, we assessed the effect of integrating medical history data in the final classification. RESULTS In a multiclass comparison between GON-confirmed eyes, GON-suspected eyes and NORMAL eyes, our model achieved 0.941 (95% confidence interval [CI], 0.936-0.946) accuracy, 0.957 (95% CI, 0.953-0.961) sensitivity, and 0.929 (95% CI, 0.923-0.935) specificity. The AUC distinguishing referrals (GON-confirmed and GON-suspected eyes) from observation was 0.992 (95% CI, 0.991-0.993). Our best model had a kappa value of 0.927, while the two experts' kappa values were 0.928 and 0.925 independently. The best 2 binary classifiers distinguishing GON-confirmed/GON-suspected eyes from NORMAL eyes obtained 0.955, 0.965 accuracy, 0.977, 0.998 sensitivity, and 0.929, 0.954 specificity, while the AUC was 0.992, 0.999 respectively. Additionally, the occlusion testing showed that our model identified the neuroretinal rim region, retinal nerve fiber layer (RNFL) defect areas (superior or inferior) as the most important parts for the discrimination of GON, which evaluated fundus images in a way similar to clinicians. Finally, the results of integration of fundus images with medical history data showed a slight improvement in sensitivity and specificity with similar AUCs. CONCLUSIONS This approach could discriminate GON with high accuracy, sensitivity, specificity, and AUC using color fundus photographs. It may provide a second opinion on the diagnosis of glaucoma to the specialist quickly, efficiently and at low cost, and assist doctors and the public in large-scale screening for glaucoma.

中文翻译：

---

基于深度学习的彩色眼底照片上青光眼视神经病变的自动检测。

目的开发基于深度残留神经网络（ResNet101）的深度学习方法，以使用彩色眼底图像自动检测青光眼性视神经病变（GON），了解模型进行预测的过程，并探索整合的效果患者的眼底图像和病史数据。方法回顾性分析2371名成年患者队列中的34279张眼底图像和相应的病史数据，并由8名青光眼专家对这些图像进行标记，其中26585张眼底图像（经GON确认的眼睛为12618张图像，经GON确认的眼睛为1114张图像）包括GON怀疑的眼睛和12,853正常眼睛的图像。我们采用10倍交叉验证策略来训练和优化我们的模型。在一个独立的测试数据集中测试了该模型，该数据集包含来自249位患者的3481张图像（1524张来自正常眼睛的图像，1442张来自经GON确认的眼睛的图像和515张来自GON怀疑的眼睛的图像）。此外，将最佳模型的性能与两位专家的结果进行了比较。计算准确性，敏感性，特异性，κ值和受体工作特征（AUC）下的面积。此外，我们对模型预测和遮挡测试进行了定性评估。最后，我们评估了将病史数据整合到最终分类中的效果。结果在经GON确认的眼睛，可疑GON的眼睛和NORMAL眼睛的多类比较中，我们的模型获得了0.941（95％置信区间[CI]，0.936-0.946）的准确度，0.957（95％CI，0.953-0.961）的敏感性，和0.929（95％CI，0.923-0。935）的特异性。观察到的AUC区分推荐人（经GON确认和可疑GON的眼睛）为0.992（95％CI，0.991-0.993）。我们最好的模型的kappa值为0.927，而两位专家的kappa值分别为0.928和0.925。从正常眼区分GON确认/ GON怀疑的眼睛的最佳2个二分类器获得的特异性为0.955、0.965、0.977、0.998和0.929、0.954，而AUC分别为0.992、0.999。此外，咬合测试表明，我们的模型将神经视网膜边缘区域，视网膜神经纤维层（RNFL）缺损区域（上或下）识别为对GON进行区分的最重要部分，并以类​​似于临床医生的方式评估了眼底图像。最后，眼底图像与病史数据整合的结果显示，与类似的AUC相比，敏感性和特异性略有改善。结论该方法可以使用彩色眼底照片以高准确度，灵敏度，特异性和AUC来区分GON。它可以为专家提供快速，有效和低成本的青光眼诊断的第二意见，并协助医生和公众进行青光眼的大规模筛查。