Year: 2025
Type: article
Abstract: Cerebrovascular segmentation in Time-of-Flight magnetic resonance angiography (TOF-MRA) is a fundamental task for assisting diagnosis, aiding in the detection and analysis of vascular structures to assist clinicians in medical decision-making. Due to the complexity of vascular structures and low background contrast, automated segmentation of brain vessels remains highly challenging. Traditional U-Net architectures struggle to effectively capture long-range and multi-scale vascular features, which limits their ability to distinguish fine structural details across slices and may lead to voxel misclassification due to convolution’s local receptive field. While CNN-Transformer hybrid models have shown promising results, we propose SWD3-Net, a novel deep learning framework incorporating a Spatiotemporal Feature Extraction module that leverages dilated convolutions to capture multi-scale contextual information across depth layers, enhancing inter-slice feature representation. In addition, we introduce a weighted binary cross-entropy loss function to address severe class imbalance. Results show SWD3-Net surpasses advanced models and the state-of-the-art deep learning medical image segmentation benchmark nnU-Net, achieving a 2% higher DSC. Our method effectively preserves vascular structures and maintains topological consistency.
Authors: Wenxin Hua, Guang Li, Guangxiang Si, Yan Li, Ning Gu
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Awards: G2293906802, G8911388731
OpenAlex ID: W4409237106