LQUnet: a vascular segmentation network based on multi-scale feature fusion and hierarchical self distillation

Vessel segmentation has important clinical significance in medical image analysis, especially in the assessment of liver disease and preoperative planning for the accurate extraction of vascular structures puts forward higher requirements. To cope with the challenges of complex vascular morphology, diverse scales and easy loss of fine vessels, this paper proposes a novel vascular segmentation network- LQUnet, which is based on supervised learning and incorporates a centreline guidance strategy to enhance the attention to the target vascular region while maintaining the ability to model the global contextual information. LQUnet introduces a gated attention mechanism and a multi-scale feature fusion module in the encoder to enhance the feature extraction capability of complex structures, and combines jump connection and branch reconstruction mechanisms in the decoder to achieve detail restoration. In addition, a composite loss function combining cross-entropy, Dice, Focal loss and hierarchical self-distillation is designed to strengthen the model's ability to learn fine-grained features. Experimental results on several publicly available datasets and self-constructed hepatic vessel datasets show that LQUnet outperforms existing methods in terms of overall segmentation accuracy and structural preservation of fine-grained vessels, and has good potential for clinical applications.