基于多层级联融合网络的糖网病检测

Detection of diabetic retinopathy based on multi-layer cascade fusion network

糖尿病视网膜病变图像的快速分析和自动识别关键在于糖网病图像特征的提取,针对糖网病特征的不明显和病灶种类多,分布、形态各异等问题,使用了OCT图像数据集代替眼底图像数据集.提出了一种基于多层级联融合网络的糖网病检测方法,可高效、准确的识别与提取高分辨率的OCT图像中每一类的病变的样本特征,运用这些特征训练softmax分类器,用于OCT图像的自动识别.在训练过程中,使用空间金字塔结构对多层的特征图处理,并将处理后的特征进行融合.结果表明该算法能有效的识别OCT图像类别,较其他的神经网络模型,具有更高的准确率和综合评价标准.

The key to rapid analysis and automatic identification of diabetic retinopathy images is the extraction of image features of glycocalyx disease.OCT image data is used for the inconspicuous characteristics of glycocalyx disease,the variety of lesions,the distribution and the different shapes.In order to replace the fundus image dataset,a method based on multi-layer cascade fusion network for detecting sugar net diseases is proposed,which efficiently and accurately identifies and extracts sample features of each type of lesion in high-resolution OCT images.These features are used to train the softmax classifier for automatic recognition of OCT images.In the training process,the spatial pyramid structure is used to process the multi-level feature maps,and the processed features are merged.The results show that the algorithm can effectively identify OCT image categories,and has higher accuracy and comprehensive evaluation criteria than other neural network models.