Weaving Attention U-net: A Novel Hybrid CNN and Attention-based Method for Organs-at-risk Segmentation in Head and Neck CT Images

Zhuangzhuang Zhang, Tianyu Zhao, Hiram Gay, Weixiong Zhang, Baozhou SunarXiv 2021 · arXiv

Introduction

In radiotherapy for head and neck cancer, the precise delineation of organs-at-risk (OARs)—such as the brainstem, optic nerves, and salivary glands—is critical. Accurate "contouring" ensures that high-dose radiation is delivered to the tumor while minimizing damage to surrounding healthy tissue. Historically, this process has been performed manually, requiring an average of 2.7 hours per patient.

While deep learning models, particularly Convolutional Neural Networks (CNNs) like U-net, have automated parts of this workflow, they face a fundamental challenge: standard convolutions excel at local feature extraction but struggle to capture long-range spatial dependencies. This limitation is particularly evident when segmenting small, low-contrast structures where global context is necessary for accurate localization.

Methodology: The WAU-net Architecture

To address these limitations, the Weaving Attention U-net (WAU-net) was developed. This novel hybrid architecture integrates the local feature-extraction capabilities of CNNs with the global context-awareness of self-attention mechanisms.

The WAU-net architecture integrates CNN blocks with axial self-attention within a nested U-net++ framework.

The WAU-net design incorporates three key components:

  1. CNN Blocks: Standard 3x3 convolutions capture low-level local features like edges and textures.
  2. Axial Self-Attention Blocks: To overcome the computational complexity of standard self-attention, WAU-net utilizes axial attention. This method performs attention along the height and width axes independently, capturing global context efficiently for high-resolution CT scans.
  3. U-net++ Backbone: The model employs a nested, dense skip-connection structure to ensure high-resolution features from the encoder are effectively fused with the decoder, preserving fine semantic details.

Experimental Results

The WAU-net was evaluated on an in-house dataset of 115 patients and the public Head and Neck 2015 Challenge dataset. The model segments ten different OARs simultaneously in a single forward pass.

Qualitative comparison of WAU-net predictions against ground truth labels across various head and neck CT slices.

Performance on Small Organs

The most significant improvement was observed in small, low-contrast organs. For the optic chiasm, WAU-net achieved a 15.6% improvement in the Dice Similarity Coefficient (DSC) over baseline models. By leveraging global landmarks (e.g., the skull base) via the attention mechanism, the model can orient itself more effectively than local-only CNNs.

Efficiency and Accuracy

WAU-net processes a full patient scan in less than two minutes, a massive improvement over manual contouring. In the in-house dataset, it achieved a DSC of 0.92 for the brain stem and 0.95 for the mandible. On the 2015 Challenge benchmark, WAU-net matched or outperformed state-of-the-art multi-stage pipelines while operating as a more efficient single-pass model.

Conclusion

WAU-net represents a significant step forward in automated OAR segmentation. By "weaving" axial attention into a dense U-net framework, the model achieves superior accuracy on challenging small structures without the need for complex multi-stage processing. Future work will focus on optimizing the computational efficiency of the dense connections and validating the model across diverse multi-institutional datasets.

Summary updated

BibTeX
@misc{zhang2021weaving,
  author = {Zhuangzhuang Zhang and Tianyu Zhao and Hiram Gay and Weixiong Zhang and Baozhou Sun},
  title = {{Weaving Attention U-net: A Novel Hybrid CNN and Attention-based Method for Organs-at-risk Segmentation in Head and Neck CT Images}},
  year = {2021},
  eprint = {2107.04847},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2107.04847},
}
Figure