Deep learning–based basilar artery wall and lumen segmentation from 1-mm MR vessel wall imaging

To investigate the morphology of the basilar artery (BA) using 1-mm magnetic resonance (MR) vessel wall imaging (VWI). This retrospective study included 36 patients who underwent intracranial 1-mm MR-VWI. The BA morphology was evaluated following a machine learning paradigm. Twenty patients (1073 cross-sectional BA images) were used to fine-tune a pre-trained deep learning model, Mask-RCNN, for BA segmentation. Six (373 cross-sectional BA images) were used for model validation and 10 (186 axial BA images) for comparison with human expert ratings. Human expert ratings were conducted in radial directions oriented at 3, 6, 9, and 12 o’clock. Agreement between human expert and machine estimation was evaluated using the intraclass correlation coefficient (ICC) and statistical significance was estimated by paired student’s t-test. BA wall segmentation was assessed using the intersection-over-union (IOU) metric. The BA exhibits a tapered shape, with the widest diameter at the beginning (3.17 ± 0.69 mm) and significantly narrowing towards the end (2.71 ± 0.55 mm) (p-value < 0.001). The deep-learning model demonstrated moderate to excellent agreement with human expert ratings (ICC: 0.72–0.83) when measuring BA diameter. However, agreement was less optimal (ICC < 0.5) when measuring artery wall thickness. For vessel wall segmentation, the model achieved a mean IOU score of 0.756 ± 0.079. This study demonstrates the effectiveness of using a 1-mm MR-VWI protocol for characterizing and evaluating the vertebrobasilar circulation. This enhanced knowledge of basilar artery shape is critical and should help neurosurgeons safely diagnose and manage posterior circulation diseases.