Carotid vessel wall segmentation on 3D black-blood MRI is a key step in plaque burden assessment and atherosclerotic lesions identification. In this study, a two-stage weakly-supervised carotid vessel wall segmentation approach was developed using limited manual delineations on 3D black-blood MR images. First, a global-local context aggregation strategy was used to identify bilateral carotid arteries robustly. Then, distances from artery center to boundaries were regressed in a polar coordinate by PolarMask to segment the vessel wall. The proposed segmentation approach outperformed Attention UNet (Quantitative score: 0.795±0.170 vs. 0.729±0.208) and showed great potential in quantitative atherosclerosis analysis.