融合ResNet结构的U-Net眼底视盘分割方法

由于亮度的相似性，带病灶眼底图像的视盘分割通常会受到亮病灶干扰。现有的视盘分割方法对正常的视网膜眼底图像具有较好的分割效果，但是在带病灶的眼底图像中表现不佳。在医学图像数据样本有限的情况下，U-Net网络能实现少样本训练生成较好的分割结果。提出一种将残差结构与U-Net网络融合的视盘分割方法。残差模块的跳跃连接能将浅层特征传递给更深一层网络，实现浅层特征的重复使用，增强了图像细节学习。将该方法在两个公开数据集Messidor和Kaggle上进行验证，在干扰较多的Kaggle数据集上，其AUC和MAP分别达到0.952 1和0.838 8，证明该方法可同时学习图像细节特征和全局结构特征，能更好地区分眼底视盘与亮病灶。

U-Net Fundus Optic Disc Segmentation Method Integrating ResNet Structure

Due to the similarity of brightness， the segmentation of the optic disc of fundus images with lesions is usually disturbed by bright lesions. Existing optic disc segmentation methods have a good segmentation effect on normal retinal fundus images， but they do not perform well in fundus images with lesions. With the premise of limited medical image data samples， the U-Net network can achieve few-sample training and generate better segmentation results. This paper proposes a method of disc segmentation that integrates the residual structure and the U-Net network. The jump connection of the residual module can transfer the shallow features to the deeper network， realize the reuse of the shallow features， and improve the learning of image detail. This article mainly verified on two public data sets of Messidor and Kaggle. On the Kaggle data set with more interference， its AUC and MAP reached 0.952 1 and 0.838 8， respectively. The method proposed in this paper realizes the process of simultaneously learning image detail features and global structural features， and can better distinguish the optic disc from bright lesions.