改进型Unet:一种高效准确的视网膜血管分割方法

Improved Unet: an efficient and accurate retinal vessel segmentation method

眼底血管的形态结构是多种疾病诊断的重要依据,但高效准确分割血管是一个巨大挑战。受多尺度卷积神经网络结构启发,将多特征提取应用到U型网络,提出改进型Unet网络。抽取眼底图像的绿通道,通过镜像、旋转、平移对训练集进行数据增强;将训练集输入到改进型Unet全卷积神经网络中进行分割;对网络模型的预测结果进行全局阈值分割得到最终结果。在DRIVE眼底数据库下实验,使用GPU分割一张565×584眼底图像仅需70ms,平均准确率高达0.9565,灵敏度、特异性也分别达到了0.7961、0.9802。实验表明算法分割准确率和效率与同类先进算法相比具有较高的水平。

The morphological structure of the fundus vessels is an important basis for the diagnosis of various diseases,but efficient and accurate segmentation of blood vessels is a huge challenge.Inspired by the structure of multi-scale convolutional neural networks,A improved U-type networks is proposed.First,the green channel of the fundus image is extracted,and the training set is enhanced by mirroring,rotating,and panning.The training set is then input into the full convolutional neural network for segmentation.Finally,the global threshold segmentation is performed on the prediction result of the network model to obtain the final result.In the DRIVE fundus database experiment,using GPU to segment a 565×584 fundus image takes only 70 ms.the average accuracy rate is up to 0.9565 and sensitivity,specificity also reached 0.7961,0.9802.Experiments show that the segmentation accuracy and efficiency of this algorithm are higher than those of advanced algorithms.