基于改进的U-Net视网膜图像的神经纤维层分割算法

Neural Fiber Layer Segmentation Algorithm Based on Improved U-Net Retinal Image

视网膜图像中的神经纤维层(Retinal Nerve Fiber Layer,RNFL)作为视网膜病变的最主要最早期的特征性部位,RNFL的分割对于糖尿病视网膜病变的评估具有重要的意义.由于视网膜图像中视网膜纤维层部分的对比度相对于背景较低,边缘不明显,分割的图像中存在断裂和难以识别的情况.提出了一种基于U-Net的RNFL分割方法,将Res path、子像素卷积层(Sub-pixel Convolution)和残差模块与原始的U-Net结合,能够更好地保留边缘信息,更加准确地分割RNFL,减少分割中断裂的情况.将提出的算法与原始的U-Net和MultiResUNet进行了比较,选取Jaccard、F1、Precision和Recall四个指标作为评价指标,结果表明本文提出的算法具有更好的分割结果,优于其他两种分割算法.

Retinal Nerve Fiber Layer in the image(Retinal Nerve Fiber Layer,RNFL) as the main characteristic of the earliest retinopathy,RNFL segmentation for the evaluation of diabetic retinopathy is of great importance.Due to the Fiber Layer part of the Retinal image contrast of low relative to the background,the edge is not obvious,the segmentation of the image of fracture and it is difficult to identify.This paper proposes a RNFL segmentation method based on U-Net,Res path,Sub-pixel Convolution layer(Sub-pixel Convolution) and residual module combined with original U-Net,can better keep the edge information,more accurate segmentation RNFL,reduce the segmentation in the fracture.This article will put forward the algorithm with the original U-Net and MultiResUNet,Jaccard,F1,Precision and Recall four indices as evaluation index,the results show that the proposed algorithm has better segmentation result,It is superior to other two segmentation algorithms.