To address the challenge of domain shift faced by retinal vessel segmentation models in cross-device applications, a multisource domain adaptive network (MSDA-Net) is proposed, incorporating structural analysis techniques derived from the precise microscopic imaging of textile engineering and fiber science. The network consists of three core modules: adaptive Source Weighting (ASW) dynamically adjusts the contribution of each source domain based on the Maximum Mean Discrepancy (MMD). Hierarchical Domain Alignment (HDA) imposes different alignment strengths at different levels of the network. Topology-Preserving Loss (TPL) is used to maintain the connectivity of vascular networks. Systematic experiments on four datasets show that in the single-source Cross-Device task (DRIVE→HRF), MSDA-Net improves F1 from 0.752 to 0.887, which is significantly better than the existing methods. In the multi-source domain joint training setting, the F1 is further improved to 0.912 when using the three-source domain (DRIVE+IOSTAR+LES), and the cross-domain performance loss is effectively controlled within 5%. The results show that the proposed method has stable and significant Cross-Device generalization ability, which provides reliable technical support for the actual deployment of retinal vessel segmentation models and high-precision fiber science imaging in both multi-center clinical environments and advanced textile engineering. This cross-domain robustness ensures that the structural intricacies of biological networks and textile manufacturing fibers are accurately captured regardless of the imaging hardware used.

retinal vessel segmentation, domain adaptation, multi-source learning, cross-device generalization, textile engineering