【摘要】 目的通过卷积神经网络UNet构建光学相干断层扫描(optical coherence tomography,OCT)图像中糖尿病性黄斑水肿的自动识别模型,并通过相关指标判断其价值。方法利用开源的OCT数据集2014_BOE_Srinivasan和OCT2017训练卷积神经网络UNet模型,并结合我院2018年1月至2019年5月的60例糖尿病性黄斑水肿患者的OCT检查影像结果共同组成数据集来验证模型。最后通过该模型的损失函数变化和精确度变化,以及绘制受试者工作特征曲线来评价模型。结果卷积神经网络UNet对单张图像的处理时间在75 ms左右。且损失函数变化图显示当模型训练到一定程度后,损失数值逐渐趋于收敛。验证集的精确度变化图显示精确度可以达到0.9左右,并且随着训练次数的不断增加,精确度逐渐趋于稳定。最后根据测试结果绘制了受试者工作特征曲线,其曲线下面积达到0.902,提示该模型具有较高诊断能力。结论利用卷积神经网络UNet可以准确快速地分割出糖尿病性黄斑水肿区域,有望辅助临床医师的诊断与治疗。更多还原

【Abstract】 Objective To establish an automatic model of diabetic macular edema(DME) in optical coherence tomography(OCT) images by convolutional neural networks UNet, and determine its value through relevant indicators.Methods The convolutional neural networks UNet model trained by using the two open sources OCT data sets 2014_BOE_Srinivasan and COT2017 in combination with the OCT images of 60 DME patients from January 2018 to May 2019 was used to validate the model. The model is evaluated by loss curve, validation accuracy curve and receiver operating characteristic(ROC) curve.Results For a single image to be segmented, the processing time was only about 75 ms. The loss curve figure showed that when the model was trained to a certain level, the loss function value tended to converge. The validation accuracy figure showed that the validation accuracy of OCT data can reach 0.9 or above, and the accuracy gradually tended to be stable as the number of training iterations increased continuously. In addition, ROC curve was drawn based on test results, and the area under curve(AUC) of ROC reached 0.902, which indicated a high diagnostic capacity.Conclusion Convolutional neural networks UNet can accurately and quickly segment the DME area, which may assist the diagnosis or treatment of clinicians.更多还原