Exploring lightweight knowledge distillation with adversarial attack for efficient coronary artery vessel segmentation
Year: 2026
Type: article
Abstract: X-ray angiography is a gold standard for diagnosing and treating coronary artery disease (CAD). Accurate segmentation of coronary artery vessels is critical in the CAD diagnostic process. However, high-performance approaches involve large-scale parameters and computational demands, leading to practical limitations in clinical scenarios. Meanwhile, existing deep learning-based methods struggle with capturing long-range semantic relationships, adapting to foreground-background interference, and preserving boundary details. To tackle those limitations, we propose Systematic Knowledge Distillation Network (SKD-Net) , a novel framework that boosts the capabilities of a lightweight student network for vessel segmentation by leveraging a powerful teacher network. Specifically, the MGA module is implemented to retain the information of complex coronary artery vessels per channel while reducing the computational overhead. The FSD module is proposed to achieve semantic-shift modeling. Concretely, the FSD module calculates the feature similarity across symmetric encoder-decoder layers, enabling the transfer of structural knowledge from a cumbersome teacher network to an untrained lightweight student network. Meanwhile, we introduce ASD module to encourage the student network to capture fine-grained pixel-wise semantic information. Extensive experiments implemented on the clinical private and public angiograms datasets show that the SKD-Net is highly efficient and superior to various knowledge distillation counterparts.
Authors: Hao Dang, Xuan Nie, Shipei Jing, Peng Li
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,353
Award: G6772131732
OpenAlex ID: W7204869262