人工智能图像识别为医学影像识别带来革命性的发展.我们团队前期基于超过18万张来自EyePACS的眼底彩照构建了辅助诊断糖尿病视网膜病变(文中简称糖网)的人工智能深度学习算法模型.在实验室条件下,该模型灵敏度和特异度分别为95.3%和79.5%的高分.本研究以Airdoc眼底病智能识别平台用户上传的糖尿病患者眼底彩照34100张为测试数据集,验证该算法在真实世界中的准确性.结果真实应用中,模型检测糖网的敏感性sensitivity(判断被测者不患病的准确度)和特异性specificity(判断被测者患病的准确度)分别是94.6%、78.4%.该人工智能模型诊断糖网的准确度已超过三甲医院眼科医师平均水平,基本达到眼底病医师水平,且工作效率远远高于人类医师,为大样本人群普查、筛查是否有糖网病提供了有力工具.%Artificial intelligence image recognition brings revolutionary development to medical image recognition. This study built a deep learning algorithm based on more than 180000 colorful photos of fundus from EyePACS for automated detection of diabetic retinopathy (DR). The test results of sensitivity and specificity achieved 95.3% and 79.5% respectively under laboratory conditions. This research tested the accuracy of the algorithm in the real world by using 34100 colorful photos of fundus uploaded by the users of Airdoc retinal disease intelligent recognition center from the internet. For detecting DR, the sensitivity and the specificity of the algorithm was 94.6% and 78.4%, respectively. The results indicated that the deep learning algorithm model had copied ophthalmologists' experience of diabetic retinopathy diagnose to an auxiliary diagnosis soft with higher efficiency similar accuracy as the human doctors. It is a potential helper for DR screening in huge people in the future.
展开▼
