Vessel Segmentation and 3D Reconstruction in Ultrasound Image Sequences Using Spatiotemporal Context Constraints and Mask Guidance

Ultrasound (US) imaging provides real-time, non-invasive, and radiation-free visualization of internal tissues, and has been widely used in vascular disease diagnosis. With the development of Robotic US Systems (RUSS), large quantities of temporal US image sequences can be efficiently acquired for vascular morphology measurement. However, automated segmentation of vessel lumens in US sequences remains challenging due to low contrast, speckle noise, and the presence of vessel-like artifacts. In response to these challenges, we propose an accurate and efficient Res-ConvLSTM-Mask Network (RLM-Net) that leverages spatial and temporal information from US sequences. The network employs residual encoding for deep vascular feature extraction, integrates a Convolutional Long Short-Term Memory (ConvLSTM) module to model temporal continuity, and introduces a Mask Propagation mechanism to suppress interference from confounding structures. In addition, a hybrid 3D reconstruction based on Marching Cubes and Poisson surface reconstruction is utilized for vascular visualization. Experiments on vessel phantom and volunteer Radial Artery (RA) datasets acquired by RUSS demonstrate that RLM-Net improves segmentation accuracy compared to existing networks, with Dice scores of 0.918 ± 0.043 and 0.796 ± 0.211 on the two volunteer RA datasets. The proposed effective approach enhances segmentation reliability and reconstruction quality, providing a robust foundation for quantitative vascular analysis in US imaging.