Exploring lightweight knowledge distillation with adversarial attack for efficient coronary artery vessel segmentation

Introduction Vessel segmentation is significant for diagnosing coronary atherosclerotic heart disease. Although deep learning has achieved remarkable progression coronary vessel segmentation, existing deep learning-based methods still suffer from huge computational costs, overfitting in small sample scenarios, and traditional loss functions that fail to balance class imbalance and vascular topological continuity. Method To address these limitations, we propose KDLight, a novel and lightweight knowledge distillation framework with adversarial attack designed for coronary artery vessel segmentation. The novel framework include the three components of adversarial attacker module, generator and knowledge distillation module. Specifically, we first embed an adversarial attack module with black-box attacker before the generator. It is capable of producing meaningful perturbations and adversarial examples throughout the training process, which is beneficial for enhancing the generalization ability and decreasing over-fitting of model. The additional knowledge distillation module consists of the teacher and student refinement sub- network. The feature-level similarity distillation module and pixel-level similarity distillation module are respectively implemented to feature extraction and vessel segmentation. Different from existing knowledge distillation paradigms, our teacher module is a segmentation network that incorporates more inputs to effectively capture noise distributions and further refine segmentation predictions during training. The acquired knowledge is subsequently transferred to the student module via feature- level similarity distillation module, empowering the student to produce improved segmentation outputs. Results Extensive experiments conducted on Dataset 1 and Dataset 2 indicate that LightKD achieves superior generalization capabilities and outperforms state-of-the-art approaches in various coronary vessel segmentation tasks. Critically, these performance improvements are statistically confirmed by consistent p -values less than 0.05. Discussion Comparative experiments with recent state-of-the-art knowledge distillation segmentation methods show that KDLight attains competitive accuracy at markedly lower FLOPs and parameter counts. Moreover, the pipeline generalizes well across multiple student backbones, confirming that it can be migrated to different lightweight encoder architectures. In future work, we will collect additional multi-center angiography data to mitigate the domain gap and further improve cross-dataset generalization.