多尺度深度CNN的糖尿病视网膜病变分类

Classification of diabetic retinopathy with multi-scale depth CNN

为对视网膜图像的DR严重程度进行更准确的分类,分析图像分辨率、深浅层特征对分类效果的影响,提出一种多尺度ResNext网络模型。使用不同分辨率眼底图像交替作为输入层数据,采用加权融合深浅层特征作为全连接层分类的信息,使用迁移学习方法对网络模型参数进行初始化,避免发生过拟合问题。基于kaggle竞赛数据的验证结果表明,相对于传统模型,该模型方法可以更准确地对DR严重程度进行分类。

To classify the severity of DR more accurately in retinal pathology images,a multi-scale ResNext network model was proposed considering the influence of image resolution and depth features on classification effect.Fundus images with different resolutions were used as input layer data alternately,and weighted fusion features of deep and shallow layers were used as classification information of all connected layers.Meanwhile,transfer learning method was used to initialize network model parameters to avoid over-fitting.Verification by using kaggle competition data shows that compared with the traditional model,the model can classify and evaluate DR severity more accurately.