作者:陈宇，徐仕豹
Authors:CHEN Yu,XU Shibao摘要:针对糖尿病视网膜病变( DR) 检测模型在下采样过程中关键信息丢失和模型鲁棒性差的问题，构建一 个PM-Net 模型( Parallel Multi-scale Network) 。在下采样过程中，利用信息增强的方式设计了多尺度最大池化和多尺度卷积模块并对 ResNet-50 改进 。进一步为了提高模型的鲁棒性，使用双分支的架构对模型进行扩展 。提出的多尺度模块使得模型在下采样的过程中获得了更加丰富的视网膜眼底图像特征，从而提高了 DR 检测的性能，同时提出的双分支模型在 DR 检测过程中用局部信息辅助全局信息保证了模型的鲁棒性 。模型在EyePACS、DDR 和私有数据集进行了实验验证 。实验结果表明: 与主流的模型相比，本模型在EyePACS 数据集上的准确率和二次加权 Kappa 分数分别提高了 2. 58% 和 1. 31% 。
Abstract:A PM-Net model ( Parallel Multi-scale Network) has been constructed to solve the problems of loss of key information and poor model robustness in the downsampling process of diabetic retinopathy ( DR) detection models. Multi-scale maximum pooling and multi-scale convolution modules have been designed and improved on ResNet-50 using information augmentation in the downsampling process. In addition，to improve the robustness of the model，the model was extended using a two-branch architecture. The proposed multi-scale module allows the model to obtain richer retinal fundus image features during downsampling，thus improving DR detection performance ， while the proposed two-branch model ensures the robustness of the model with local information supplementing global information during DR detection. The model was experimentally validated on EyePACS ，DDR , and private datasets. Experimental results show that the model accuracy and quadratic weighted kappa score on the EyePACS dataset are improved by 2. 58% and 1. 31% respectively compared to mainstream models.


PDF全文下载地址:

可免费Download/下载PDF全文