Lightweight U-Net++ with Hybrid Swin-Convolutional Attention for Efficient CT Metal Artifact Reduction
Frontiers in Computing and Intelligent Systems · 30.04.2026 · 16 (2) · 18-22
DOI: 10.54097/hcg6tr24 ↗Kurz gefasst
Fachartikel in der Zeitschrift Frontiers in Computing and Intelligent Systems, online erschienen am 30.04.2026. Verfasst von Shanshan Wang und 4 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Metal artifacts from high-density implants severely degrade CT images, hindering accurate diagnosis and downstream tasks. Existing Metal Artifact Reduction (MAR) methods face a critical trade-off: CNNs struggle with global artifacts due to limited receptive fields, while Transformers incur prohibitive computational costs unsuitable for real-time clinical use. To address this, we propose U-Net++(SwinLite), an efficient, lightweight MAR architecture. By integrating a nested dense cascade skip-connection mechanism and a Hybrid Attention Module (coupling Swin and convolutional attention), our model effectively captures both local details and long-range dependencies. Using a dynamic channel adjustment strategy, U-Net++(SwinLite) achieves a Mean Absolute Error (MAE) of 9.6006 with only 3.05M parameters —reducing model size by over 95% compared to pure Transformer models like Restormer-U. Extensive experiments demonstrate our method achieves an optimal balance between structural preservation and computational efficiency.
Zitieren
APA
Wang, S., Zang, A., Zhang, H., Lu, P., & Ma, J. (2026). Lightweight U-Net++ with Hybrid Swin-Convolutional Attention for Efficient CT Metal Artifact Reduction. Frontiers in Computing and Intelligent Systems, 16(2), 18-22. https://doi.org/10.54097/hcg6tr24
Vancouver
Wang S, Zang A, Zhang H, Lu P, Ma J. Lightweight U-Net++ with Hybrid Swin-Convolutional Attention for Efficient CT Metal Artifact Reduction. Frontiers in Computing and Intelligent Systems. 2026;16(2):18-22. doi:10.54097/hcg6tr24
BibTeX
@article{Wang2026,
author = {Wang, Shanshan and Zang, Ao and Zhang, Hao and Lu, Peng and Ma, Jiang},
title = {Lightweight U-Net++ with Hybrid Swin-Convolutional Attention for Efficient CT Metal Artifact Reduction},
journal = {Frontiers in Computing and Intelligent Systems},
year = {2026},
volume = {16},
number = {2},
pages = {18--22},
doi = {10.54097/hcg6tr24}
}
EUVIMED