双线性非局部特征结合中继监督网络用于视网膜血管分割

对眼底图像中的视网膜血管精准分割是检测多种疾病的关键技术，在相关疾病自动筛查系统中发挥着重要的作用。针对现存方法追求分割精度时忽略对算法复杂性的考虑，导致在资源受限的医疗设备上部署困难的问题。本文通过进一步合理减少卷积层的特征通道数量来轻量化分割网络并提出了BNIS-Net。该网络采用多尺度图像作为输入融合到编码过程中，使得不同感受野之间建立良好的联系，并提出一种双线性非局部模块来增强相关上下文信息的捕捉能力。最后，在解码过程中采用中继监督的策略，为解码部分各级输出提供监督来约束网络的学习，这样可以有效改善收敛行为使浅层部分得到充分训练。BNIS-Net以0.41 M的参数量在DRIVE、STARE和CHASE_DB1 3个公开数据集上分别取得了81.02%、81.07%、78.15%的DSC值和0.983 3、0.986 1、0.985 9的AUC值。通过大量对比实验和消融研究证明，该方法能够更好地分割血管的边缘细节。

Bilinear non-local features combined with intermediate supervision network for retinal vessel segmentation

Abstract: The accurate segmentation of retinal vessels in fundus images is of significance for the detection of various diseases and plays an important role in automated screening system for associated diseases. To address the problem that existing methods neglect to consider the complexity of the algorithm when pursuing segmentation accuracy, which leads to difficulties in deployment on resource-constrained medical devices, the number of feature channels in the convolutional layer is further reduced to lighten the segmentation network, and a bilinear non-local intermediate supervision network (BNIS-Net) is proposed. In BNIS-Net, the multi-scale images are taken as input and fused into the coding for establishing good connections between different receptive fields, and a bilinear non-local module is added to enhance the capture of relevant contextual information. During the decoding, an intermediate supervision strategy is adopted to constrain the learning of the network by providing supervision to the output of the decoding at all levels, which can effectively improve the BNIS-Net uses a parameter of 0.41 M on 3 public data sets of DRIVE, START and CHASE, and achieves DSC values of 81.02%, 81.07% and 78.15%, and AUC values of 0.983 3, 0.986 1 and 0.985 9, respectively. It was demonstrated by numerous comparative experiments and ablation studies that the method can better segment the edge details of vessels.