双路径编码和跨级解耦的视网膜血管分割

打开文本图片集

关键词：视网膜血管；图像分割；多向差分卷积；动态特征融合；注意力机制中图分类号：TP394.1；TH691.9 文献标识码：Adoi：10.37188/OPE.20253314.2242 CSTR：32169.14.OPE.20253314.2242

Dual-path encoding and cross-level decoupling for retinal vessel segmentation

ZHANG Rongrong1， LU Xiaoq 1，2* ，LI Jing¹，GU Yul，LIU Chuanqiang1

（1.School ofDigital and Intelligence Industry，Inner Mongolia University ofScience and Technology，Baotou O14010， China; 2.School ofInformation Engineering，InnerMongolia University of Technology，Hohhot O1Oo51，China） * Corresponding author，E-mail： lxiaoqi@imut. edu. cn

Abstract：Retinal vessel segmentation is a critical foundation for ophthalmic disease diagnosis.However， existing methods have issues such as missed detection of thin vessels，interference from pathological regions，and feature entanglement. Therefore，this paper proposed a Dual-path Feature Extraction and Cro-level Feature Decoupling Network.First，in the encoder，a Multi-directional Diferential Residual Block extracted vessel-edge features across multiple directions to capture subtle vessel structures.Next， aCollaborative Attention Fusion Block dynamically integrated the complementary convolutional features of the double-branch path，enhancing the aggregation of encoded information.Finally，for the skip connections between the two U-Net branches，a Channel Interaction Disentanglement Block was introduced to decouple cross-level features， strengthen feature representations，and resolve the feature-confusion problem inherent in traditional U-Net architectures. The proposed method was extensively validated on four public datasets—DRIVE，CHASEDB1，STARE，and IOSTAR—achieving F1-scores of 82.47% ， 80.71% ， 81.44% ，and 82.01% ，and sensitivities of 80.96% ， 80.23% ， 74.69% ，and 76.92% ， respectively. The F1 scores of the LadderNet algorithm are 81.66% ，80. 16% ， 80.92% and 79.69% ， and the sensitivity is 77.06%，78.88%，73.64% and 71.24% ，respectively. Compared to state-of-the-art methods，our approach shows good robustness and superior segmentation performance in the task of retinal vascular segmentation.

Key words：retinal vessel; image segmentation；multi-directional diferential convolution；dynamic feature fusion；attention mechanism

1引言

视网膜血管是人眼的重要组成部分，其血管分布形式可以反映出眼底的健康状况，不同的病变会导致血管分布的改变，因此通过分析眼底血管的分布情况，可以对相关的眼底疾病（如，糖尿病视网膜病变，高血压和其他心血管疾病，年龄相关的黄斑变性等）进行诊断[1]。（剩余22504字）