基于深度迁移学习的糖尿病视网膜病变的检测

Identifying diabetic retinopathy based on deep transfer learning

为了自动检测视网膜眼底图像中的糖尿病视网膜病变(DR),缩减眼科医生工作量,提供视网膜疾病检测和诊断的辅助工具,提出了基于Inception-v3模型的深度迁移学习方法对DR图像进行自动检测。使用ImageNet大数据集预先训练过的Inception-v3模型,将之前传递层参数固定,采用不断微调的方法,通过自行收集的数据集对模型的最后一个完全连接层进行重新训练得到新的分类器。实验结果表明,所提出的方法无需指定病变的特征就能够获得高精度预测和高可靠性检测。除了帮助眼科医生作出诊断决定之外,还可以基于视网膜眼底图像帮助自动筛查早期DR。

In order to identify diabetic retinopathy(DR)in retinal fundus images automatically,to reduce the workload of ophthalmologists,and to develop an assistant tool in detecting and diagnosing retinal diseases,automated detection of DR images which uses deep transfer learning approach based on the Inception-v3 model is proposed.In the Inception-v3 model trained by ImageNet datasets,the parameters of the previous layers were fixed while the last fully-connected layer of the model was retrained by fine-tuning on the dataset collected by ourselves.Experimental results manifested the performance of the proposed approach providing better predictions and highly reliable detection without specifying lesion-based features,and it could help make automated screening for early DR based on retinal fundus images in addition to assisting ophthalmologists in making a referral decision.