Claw U-Net: A Unet-based Network with Deep Feature Concatenation for Scleral Blood Vessel Segmentation

Chang Yao, Jingyu Tang, Menghan Hu, Yue Wu, Wenyi Guo, Qingli Li, Xiao-Ping ZhangarXiv 2020 · arXiv

Abstract

Automated segmentation of scleral blood vessels is a critical task for diagnosing and treating Sturge-Weber syndrome (SWS). We present Claw U-Net, a specialized neural network architecture designed to handle the high density and scale variation of scleral vasculature. By repeatedly "clawing" high-level semantic features from the deepest network layer and feeding them into every decoding stage, the model achieves state-of-the-art accuracy. Experimental results on a newly established surgical dataset show that Claw U-Net significantly outperforms standard U-Net and U-Net++ architectures in both quantitative metrics and qualitative detail preservation.

Introduction

Sturge-Weber syndrome (SWS) is a serious vascular condition that can lead to glaucoma and blindness. The pattern and density of blood vessels on the sclera are key indicators of disease severity and surgical prognosis. However, scleral vessel segmentation is challenging due to high vessel density, significant scale variation, and lower image quality typical of surgical environments. Currently, clinical assessment relies on manual grouping based on experience, which lacks standardization. There is a pressing need for an automated, real-time tool to map these vessels accurately.

Methods

Claw U-Net builds upon the classic U-Net and U-Net++ architectures, introducing three specific enhancements to bridge the semantic gap and improve fine-grained segmentation:

  1. Deep Feature Concatenation (The "Claw" Structure): Unlike standard U-Nets where the deepest features are only used at the bottleneck, Claw U-Net continuously upsamples these bottom-layer feature maps and concatenates them with every layer of the decoder. This ensures that robust high-level semantic information is preserved throughout the reconstruction process.
  2. Residual Learning: The encoder utilizes a ResNet-34 backbone. Residual blocks allow for a deeper network without vanishing gradient issues, enabling the learning of more complex vascular features.
  3. Triple-Input Attention Mechanism: A novel attention gate is integrated into the decoding stage. It processes three inputs: encoder features, previous decoder features, and upsampled features from the deepest layer. This helps the model focus on small vessel boundaries and fine-grained regions.
Architecture of Claw U-Net showing the deep feature concatenation (claw) and attention gates.

Datasets and Experimental Setup

The study utilizes a new scleral blood vessel dataset comprising 51 high-resolution (512 × 512) images from 51 patients, captured during surgery. Ground truth labels were provided by experienced ophthalmologists. Despite its modest size, the dataset represents a high-quality, clinically relevant challenge for automated segmentation.

Results and Discussion

The ResClaw UNet + attention variant achieved superior performance across all key metrics:

  • MIoU: 80.57% (vs. 79.50% for standard U-Net).
  • Dice Coefficient: 88.58%, indicating high overlap with manual labels.
  • Average Hausdorff Distance: Reduced to 11.99, signifying more accurate boundary prediction.

Qualitative analysis confirms that Claw U-Net is particularly effective at preserving tiny, "broken" vascular segments that are often missed by traditional models.

Qualitative comparison: (a) Input image, (b) Ground truth mask, (c) and (d) Model predictions showing detailed vessel recovery.

Conclusion

Claw U-Net demonstrates that re-injecting deep semantic features into the decoder significantly improves the segmentation of complex, fine-grained structures like scleral vessels. Future work will focus on expanding the dataset to improve generalization across different lighting and hardware conditions, and optimizing the network for real-time surgical assistance.

Summary updated

BibTeX
@misc{yao2020claw,
  author = {Chang Yao and Jingyu Tang and Menghan Hu and Yue Wu and Wenyi Guo and Qingli Li and Xiao-Ping Zhang},
  title = {{Claw U-Net: A Unet-based Network with Deep Feature Concatenation for Scleral Blood Vessel Segmentation}},
  year = {2020},
  eprint = {2010.10163},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2010.10163},
}
Figure