Carotid Artery Segmentation from 3D Ultrasound Images Using Contextual YOLO and Adaptive Prompted SAM 2

Stroke is a leading cause of mortality and disability worldwide, with carotid atherosclerosis being a major contributor to ischemic stroke. Accurate segmentation of the lumen-intima boundaries (LIB) and media-adventitia boundaries (MAB) in 3D ultrasound (3DUS) images is important for quantifying biomarkers such as vessel wall volume (VWV), which are essential for stroke risk assessment and therapy evaluation. However, manual segmentation is time-consuming, prone to observer variability, and challenged by obscured vessel wall boundaries caused by artifacts in ultrasound imaging. To address these limitations, we propose a novel automated segmentation framework incorporating contextual information and iterative refinement. The framework integrates three key components: (1) a Contextual YOLO module that leverages information from neighboring slices to generate bounding box prompts, (2) a frozen SAM 2 model that performs segmentation using these prompts without requiring additional training, and (3) an adaptive prompt adjustment mechanism (APAM) that refines bounding boxes and segmentation masks based on inter-slice consistency. One advantage of our framework is the interactive capability of SAM 2, which enables clinicians to refine segmentation results by providing additional prompts in clinical practice. Experimental results demonstrate the effectiveness of our segmentation framework, offering a scalable solution for 3DUS-based stroke risk assessment and therapy evaluation.