基于Shuffle-Unet的视网膜血管分割研究

Research on retinal vessel segmentation based on Shuffle-Unet

针对传统视网膜血管分割算法检测速度慢,难以应用于实时医疗辅助诊断系统的问题,提出一种轻量型的基于Shuffle-Unet的视网膜血管分割模型。对轻量级模型ShuffleNetV2进行结构剪枝,剪除ShuffleNetV2结构上最后一层卷积层、全局池化层和全连接层,简化模型结构;将剪枝后的ShuffleNetV2作为模型的主干提取网络,降低模型的计算复杂度,提高模型的运行速度;使用随机通道分离操作模块搭建上采样模型结构,增强网络特征传递能力;使用注意力机制模块将模型的第一层特征层输出和上采样层相融合,分别从通道和空间两个维度上增强模型对有效特征的提取。通过DRIVE、CHASE_DB1两个公开数据集与其他视网膜血管分割算法进行对比,有效的证明了Shuffle-Unet模型具有高分割精度和高检测速度的特点。

Aiming at the slow detection speed of traditional retinal blood vessel segmentation algorithm, it is difficult to apply to real-time medical aided diagnosis system, a lightweight retinal blood vessel segmentation model based on Shuffle-Unet is proposed. In order to simplify the model structure, the lightweight model ShuffleNetV2 is structurally pruned, and the last convolutional layer, global pooling layer and fully connected layer on the ShuffleNetV2 structure are pruned;In order to reduce the computational complexity of the model and improve the running speed of the model, the pruned ShuffleNetV2 is used as the backbone extraction network of the model;Use the random channel separation operation module to build an upsampling model structure to enhance the network feature transfer capability;The attention mechanism module is used to fuse the output of the first feature layer and the upsampling layer of the model to enhance the extraction of effective features from the two dimensions of the channel and the space. By comparing the two public datasets DRIVE and CHASE_DB1 with other retinal blood vessel segmentation algorithms, it effectively proves that the Shuffle-Unet model has the characteristics of high segmentation accuracy and high detection speed.