基于多任务学习的青光眼智能诊断

Intelligent Diagnosis of Glaucoma Based on Multi-Task Learning

为了提高青光眼检测的准确率,降低青光眼的危害,本文提出一种基于多任务学习的青光眼智能诊断方法,将U-Net网络和VGG16网络结合,U-Net网络和VGG16网络共用U-Net网络的编码器部分,通过U-Net网络得到杯盘比(cup-to-disc ratio,CDR),并且将CDR作为眼底图像的特征之一输入VGG16网络,实现眼底图像的青光眼分类。实验使用REFUGE挑战数据集进行验证,网络模型在训练后得到的工作特性曲线下面积为0.9788,且视盘和视杯的分割准确率分别达到0.8745和0.9624,对比其他使用相同数据集的方法,本方法具有更高的青光眼分类准确率。

In order to enhance the accuracy of glaucoma detection and mitigate the risks associated with glaucoma,in this article we propose an intelligent diagnostic method for glaucoma based on multi-task learning.Our proposed method combines the U-Net and VGG16 networks,with the encoder part of the U-Net network being shared by both networks.By utilizing the U-Net network,the cup-to-disc ratio(CDR)is obtained from retinal images,and this CDR is used as one of the features input into the VGG16 network to achieve glaucoma classification for the retinal images.The proposed method was validated using the REFUGE challenge datasets.After training the network model,the area under the receiver operating characteristic curve(AUC)was measured to be 0.9788.Moreover,the segmentation accuracy for the optic disc and optic cup was found to be 0.8745 and 0.9624,respectively.In comparison to other methods using the same datasets,the proposed method in this article demonstrates higher accuracy in glaucoma classification.