探讨基于人工智能深度学习结合视盘改变分类多任务模型对青光眼的检测能力。

收集2017年8月至2018年8月来自29个省直辖市自治区的医疗机构的共计21 969例(32 548只眼)患者的临床资料。其中，男性12 762例(18 745只眼)，女性9207例(13 803只眼)；年龄17~75岁，平均年龄(57.1±7.9)岁。由两级人工阅片小组根据眼底图像诊断并标注为正常眼与青光眼，并将其划分为训练集、同源测试集及非同源测试集。使用基于Python 3.6和Pytorch 0.4的腾讯觅影眼底照片辅助诊断系统，以NVIDIA Tesla P40作为运行芯片完成模型的训练，以ResNet 34作为骨干网络的多任务卷积神经网络模型，以青光眼分类任务为主任务，视盘萎缩任务为辅助任务，并选取在调优集上表现最好的模型作为最终模型。此外，利用模型类别激活图对模型做出预测的特征区域进行解释。采用敏感度、特异度及工作特性曲线下面积对其进行预测和性能评价。

模型在三个测试集中诊断青光眼的灵敏度分别为95.9%、95.4%及95.7%；特异度分别为97.7%、91.6%及92.2%；工作特性曲线下面积分别为0.993、0.968及0.974。在引入视盘萎缩多任务后，敏感度和特异度得到了有效提升。

基于深度学习结合视盘改变分类的多任务模型对于青光眼的检测具有较高的准确率。同时，通过可解释性实验有效探索和分析了本模型对于高度近视眼等具有与青光眼相似眼底特征的鉴别能力。

The aim of this study was to investigate the detection ability of multi-task model based on deep learning combined with optic disc change classification in glaucomatous optic neuropathy.

21 969 participants (32 548 eyes) were included from medical institutions of 29 Provinces, municipalities and autonomous regions from August 2017 to August 2018. There were 12 762 male and 9207 female, and the age range was 17 to 75 with the average age (57.1±7.9) years-old. The two-level group diagnosed and labeled the fundus images as normal eyes and glaucomatous eyes, and divided them into training set, homologous testing set and two non-homologous testing sets. Our research was based on Tencent Miying Fundus Photo Auxiliary Diagnosis System using Python 3.6 and Pytorch 0.4. We trained our model with NVIDIA Tesla P40 as running chip. Using the multi-task convolution neural network model based on ResNet 34, with the main task of glaucoma classification and the auxiliary task of optic disc atrophy, the model with the best performance in the optimization set was selected as the final model. Moreover, class activation map was used to explain the feature areas of the model. The sensitivity, specificity and area under the curve are used to predict and evaluate its performance.

The sensitivity of the model was95.9%, 95.4% and 95.7%, respectively; specificity of that was 97.7%, 91.6% and 92.2%, respectively; area under the curve of that was 0.993, 0.968 and 0.974, respectively. The sensitivity and specificity of the model were effectively improved after the introduction of multi-task model of optic disc atrophy.

The multi-task model based on deep learning combined with optic disc change classification has a high accuracy for the detection of glaucomatous optic neuropathy. Moreover, the ability of this model to distinguish high myopia from glaucoma was explored and analyzed by interpretable experiment.