Wavelet-enhanced spatiotemporal connectivity-preserving network for intracranial artery segmentation in DSA sequences

Abstract Accurate intracranial artery segmentation from Digital Subtraction Angiography is critical for the diagnosis and interventional treatment of cerebrovascular diseases. However, traditional methods struggle to capture dynamic spatiotemporal dependencies, often leading to vascular discontinuity and the loss of fine distal vessels due to background interference and resolution loss. In this study, we propose the Wavelet-Enhanced Spatiotemporal Connectivity-Preserving Network (WESCP-Net), a novel framework designed to synergize physical priors with frequency-domain feature extraction. Specifically, we introduce a Physically-Guided Spatiotemporal Enhancement module that explicitly exploits hemodynamic flow variance to differentiate active vascular signals from static artifacts. To address the loss of high-frequency spatial details in standard downsampling, we incorporate a Wavelet-Integrated Encoder and a Topology-Aware Reconstruction module, which utilize discrete wavelet transforms to preserve sharp vessel boundaries and restore structural connectivity. Experimental results on the DIAS dataset demonstrate that WESCP-Net achieves state-of-the-art performance, yielding a Dice Similarity Coefficient of 0.7982 and an Intersection over Union score of 0.6422. Notably, its connectivity-preserving mechanism achieves a clDice metric of 0.7135, improving the continuity of vascular terminals. WESCP-Net provides a robust technological paradigm for precise cerebrovascular segmentation, facilitating reliable surgical navigation and quantitative diagnosis.