EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Scale-Aware Micro-Vasculature Segmentation Via a ConvNeXt-Based Dual-Head Network With Skeleton Topology Regularization

Ram Kumar Yadav, Avishek Nandi, Akhilesh Kumar Sharma, Usha Choudhary

IEEE Access · 2026 · 14 · 73975-73987

DOI: 10.1109/access.2026.3692746 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Ram Kumar Yadav und 3 weiteren Autor:innen. Die Arbeit nennt 27 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Microvasculature segmentation from high-resolution medical imaging is a challenging task because of the sparse and complex nature of the vascular patterns. Conventional deep learning architectures often fail to capture the topological connectivity and the fine details of the capillaries. In this work, we propose a novel deep learning pipeline that overcomes these challenges by synergistically combining a state-of-the-art network architecture, topology-aware learning, and content-adaptive data sampling. Our proposed model is based on a ConvNeXt-Tiny backbone network architecture cast within a U-Net framework, complemented by a dual-head decoder network that simultaneously predicts the vessel masks and their corresponding skeletal centerlines. The multi-task learning framework acts as a strong regularizer, which forces the network to learn the underlying connectivity of the vascular structure. To counter the severe class imbalance problem that is inherent in vessel segmentation tasks, we adopt a “force-feed” sampling approach that ensures the training data is predominantly composed of image patches with vessels. Moreover, we incorporate a Tversky loss function that gives higher weights to false negative errors, thus promoting the detection of small and typically under-detected capillaries. Our proposed model is tested on a challenging kidney microvasculature segmentation dataset, and it achieves state-of-the-art results. Our model achieves an F1-Score of 0.850, which is significantly better than the existing state-of-the-art architectures such as Swin-UNet (0.727) and ResNet34-UNet (0.728). Our model also passes all qualitative and quantitative tests with high robustness, accuracy of 0.999, and near-perfect AUC of 0.997, thus setting a new gold standard for microvasculature segmentation.

Zitieren

APA

Yadav, R. K., Nandi, A., Sharma, A. K., & Choudhary, U. (2026). Scale-Aware Micro-Vasculature Segmentation Via a ConvNeXt-Based Dual-Head Network With Skeleton Topology Regularization. IEEE Access, 14, 73975-73987. https://doi.org/10.1109/access.2026.3692746

Vancouver

Yadav RK, Nandi A, Sharma AK, Choudhary U. Scale-Aware Micro-Vasculature Segmentation Via a ConvNeXt-Based Dual-Head Network With Skeleton Topology Regularization. IEEE Access. 2026;14:73975-73987. doi:10.1109/access.2026.3692746

BibTeX
@article{Yadav2026,
  author = {Yadav, Ram Kumar and Nandi, Avishek and Sharma, Akhilesh Kumar and Choudhary, Usha},
  title = {Scale-Aware Micro-Vasculature Segmentation Via a ConvNeXt-Based Dual-Head Network With Skeleton Topology Regularization},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {73975--73987},
  doi = {10.1109/access.2026.3692746}
}
Kostenlos starten →