基于MF-UNet网络的视网膜血管分割

Retinal vessel segmentation based on MF-UNet network

现有的视网膜血管分割算法存在特征提取能力不足和分割效率低等问题。针对该问题,对UNet网络进行改进,提出一种基于多尺度特征提取的U型网络(Multi-scale feature extraction based on UNet, MF-UNet)。该算法在编码和解码部分构建反卷积分割模块替代传统卷积块,使网络保留更多的血管细节信息。之后,在编码和解码中间连接部引入混合池化(Mix Pooling Moudle, MPM)和模板卷积(Template convolution, TConv),提升网络对多尺度特征的提取能力,从而提升血管的分割质量和分割效率。在两个眼底数据库DRIVE和STARE上进行实验验证,结果表明,MF-UNet算法在准确性、灵敏度、特异性和AUC表现优异,更优于UNet与其他视网膜血管分割算法。

The existing retinal vessel segmentation algorithms have problems involving insufficient feature extraction and low segmentation efficiency. To solve this problem, the UNet network is improved and a Multi-scale feature extraction, based on UNet(MF-UNet), is proposed. This algorithm constructs a deconvolutional segmentation module to replace the traditional volume block in the encoding and decoding sections, so that more detailed information about the vessels is retained by the network. Afterwards, mix pooling module(MPM) and template convolution(TConv) are introduced in the middle connection between encoding and decoding, which can enhance the network’s extraction capability of multi-scale features, thus improving the segmentation quality and segmentation efficiency of vessels. The experimental validation was performed on two fundus databases, namely DRIVE and STARE. The results show that the MF-UNet algorithm outperforms UNet and other retinal vessel segmentation algorithms in terms of accuracy, sensitivity, specificity and AUC.