基于短距离跳跃连接的U2-Net+医学图像语义分割

U2⁃Net+medical image semantic segmentation based on short⁃range jump connection

医学图像分割是保障发展智慧医疗系统的先决条件之一。由于原U2-Net+网络的跳跃连接只关注同分辨率所提取的特征,所以在设计时借鉴FR-UNet网络加入中间层,接收深层的上下文信息与浅层提取的高分辨率特征进行整合;并在中间层的下采样使用非对称空洞空间卷积金字塔代替,增加网络模型训练时对边缘信息的关注,并在结构最后加入阈值增强模块,加强对细小特征边缘的识别与分割;同时加入到上采样中,帮助网络更好地提取多尺度特征,增加上下文语义关联。根据正负样本不均衡和难易不同的问题设计了组合的损失函数来监督网络优化。实验结果表明,所提算法在DRIVE、STARE+CHASE_DB1数据集上的F1分数分别提高了1.8%与4.2%,在ISIC2018数据集上的DSC分数提高了2.3%。对分割结果进行可视化后表明,该网络在样本较小的情况下可以充分提取到更加精确的边缘信息和细小的特征信息,提高语义分割的效果,所提算法在医学图像语义分割任务上有更好的表现。

Medical image segmentation is one of the necessary prerequisites to guarantee the development of an intelligent medical system.Because the jump connection of the original U2⁃Net+network only focuses on the features extracted at the same resolution,an intermediate layer is added in the design by taking the FR⁃UNet network as reference.The contextual information from the deeper layer is received by the intermediate layer to integrate with the high⁃resolution features extracted from the shallower layer.In the down⁃sampling of the intermediate layer,a convolutional pyramid with asymmetric atrous space is used to increase the attention to edge information during network model training.A threshold value enhancement module is added at the end of the structure to strengthen the identification and segmentation of the edges with fine features.It is also added to the up⁃sampling to help the network extract multi⁃scale features better and increase contextual semantic associations.A combined loss function is designed to supervise network optimization according to the imbalance between the positive and negative samples and the different levels of difficulties.The experimental results show that the proposed algorithm improves the F1⁃score by 1.8%and 4.2%on the datasets of DRIVE and STARE+CHASE_DB1,respectively,and improves the DSC score by 2.3%on the dataset ISIC2018.Visualization of the segmentation results shows that the present network can fully extract more accurate edge information and fine feature information to improve the semantic segmentation in the case of smaller samples,so the proposed algorithm has a better performance on the task of semantic segmentation of medical images.