STHP-Net:A Spatio-Temporal Collaborative High-Frequency Purification Network for Intracranial Artery Segmentation in DSA Sequences

Digital subtraction angiography (DSA) is regarded as the gold standard for the diagnosis of cerebrovascular diseases owing to its high spatial resolution and dynamic blood flow imaging capability. Accurate segmentation of intracranial arteries is critical for the diagnosis, treatment, and prognosis of ischemic stroke. Nevertheless, existing methods still suffer from several challenges: traditional convolutional attention mechanisms are prone to missing small vessels due to inherent limitations in modeling local features, and current noise suppression strategies can hardly maintain a balance between preserving fine vascular details and eliminating noise at the feature level. To address these issues, we propose a Spatio-Temporal Collaborative High-Frequency Purification Network (STHP-Net). The network first exploits cross-frame complementary information in DSA sequences to maintain the integrity of vascular structures. We then design a Spatial-Channel Combined Attention Module (SCAM) to enhance the weak feature representation of thin vessels. Based on this, a High-Frequency Adaptive Purification Module (HAPM) is introduced to achieve accurate decoupling of vascular details and high-frequency noise in the feature space. Finally, a full-resolution decoding network is employed to generate the segmentation result.Experiments on the public DIAS dataset show that the Dice and clDice coefficient of STHP-Net reach 0.7909 and 0.7386, respectively, significantly outperforming state-of-the-art methods in terms of small vessel detection rate and vascular topological connectivity.