近视性黄斑病变(myopic maculopathy，MM)是近视最常见的并发症，也是影响病理性近视视功能下降的主要原因。目前，MM的分类系统尚不能完全解释患者黄斑部发生的多种变化，迫切需要一个全面、统一的分类系统来协助沟通和比较临床试验以及国际多中心研究的结果。随着眼底成像技术的发展与应用，最新的近视性黄斑病变分类，即ATN分类系统[萎缩(A)、牵拉(T)、和新生血管(N)]结合眼底照片与光学相干断层扫描(optical coherence tomography，OCT)图片，把黄斑病变分为3类，每一类又根据其严重程度进行分级，这对MM的诊断和治疗提供了较大的临床价值。

Myopic maculopathy (MM) is the most common complication of myopia, which is also the main cause of poor visual function in pathologic myopia. Presently, the classification system of MM cannot properly explain the numerous changes that occur in the patient’s macula. Therefore, a comprehensive and unified classification system is urgently needed to facilitate in communicating and comparing the results of clinical trials and international multicenter studies. With the development and application of fundus imaging technology, the latest classification of MM, namely ATN (atrophy, A; traction, T; neovascularization, N) classificationsystem, which combines fundus photographs and optical coherence tomography (OCT) images, classifies macular lesions into 3 categories according to its severity, generating greater clinical value for the diagnosis and treatment of MM.

目的：探索基于眼底彩照和人工智能构建冠心病智能诊断系统的可行性。方法：于2013—2014年收集广东省人民医院530例患者共2117张眼底彩照，其中冠心病217例共909张眼底彩照。根据患者有无冠心病的情况进行标记，使用Inception-V3深度卷积神经网络训练人工智能模型，随后使用验证数据判断模型的准确率。计算深度卷积网络模型的准确性、一致率、敏感性、特异性和受试者工作特性曲线下面积(area under the curve，AUC)。结果：在2117张眼底彩照中，1903张用于模型训练，214张用于模型的性能评估。在测试集中，该算法的准确性为98.1%，一致率为98.6%，敏感性为100.0%，特异性为96.7%，AUC为0.988(95%CI：0.974~1.000)。结论：眼底彩照联合人工智能技术可精准判定冠心病，该模型具备较高的敏感性和特异性，但须进一步增加样本量，使用大样本量数据验证该模型，排除过拟合的可能性。

Objective: To explore the feasibility of developing a deep learning algorithm for detecting coronary heart diseases based on fundus color photography and artificial intelligence (AI). Methods: A total of 2 117 fundus  color photographs were taken from 530 patients in Guangdong Provincial People’s Hospital from 2013 to 2014,including 909 fundus color photographs from 217 patients with coronary heart disease (CHD). According to whether the patient had coronary heart disease or not, the Inception-V3 depth convolution neural network was used to train the deep learning model, and then the validation data were used to judge the accuracy of the model. The accuracy, consistency rate, sensitivity and specificity of the deep convolution network model and the area under the working characteristic curve (AUC) were calculated. Results: Among the 2 117 fundus color photographs, 1 903 were used for model training, and 214 were used to test the accuracy of the model. In the test dataset, the accuracy of the algorithm was 98.1%, the consistency rate was 98.6%, the sensitivity was 100.0%, and the specificity was 96.7%. The AUC was 0.988 (95% CI, 0.974–1.000). Conclusion: The combination of fundus color photography and artificial intelligence can achieve the accurate diagnosis of the coronary heart disease, and the model has high sensitivity and specificity. However, future studies are warranted to validate our model and exclude the possibility of over-fitting.