VB-ROI-focused segmentation-to-morphometry workflow for vertebrobasilar dolichoectasia assessment on MRA

Vertebrobasilar dolichoectasia (VBD) is closely associated with posterior circulation infarction; however, reliable MRA-based morphometric quantification is hindered by the challenges of delineating the basilar artery (BA) and bilateral vertebral arteries (R-VA and L-VA). These vessels are characterized by their thin, elongated structure, spatial localization, and high sensitivity to boundary deviations, endpoint instability, and centerline discontinuity, which render generic full-image vessel segmentation suboptimal for VBD-oriented morphometric analysis.

In this study, we address this clinical-computational mismatch using a vertebrobasilar region-of-interest (VB-ROI)-focused segmentation-to-morphometry workflow. Initially, a full-image nnU-Net is employed for coarse anatomical localization, from which a prediction-derived VB-ROI is generated. Subsequently, a second nnU-Net performs fine segmentation of the BA/R-VA/L-VA within this anatomically focused field of view, and the restored full-size mask is utilized for segmentation evaluation and downstream morphometric quantification.

Compared to the direct full-image nnU-Net, the proposed two-stage VB-ROI workflow improved the mean Dice coefficient from 0.7974 to 0.9681, the mean Intersection over Union (IoU) from 0.7145 to 0.9412, and the mean centerline Dice from 0.7398 to 0.9206, while reducing the Hausdorff Distance at 95th percentile (HD95) from 8.6245 mm to 1.1709 mm. The prediction-derived VB-ROI demonstrated performance comparable to the ground-truth-based ROI setting without necessitating test-time ground truth, thus supporting the robustness of the localization strategy. Morphometric analysis indicated strong agreement in diameter-related measurements, while length, displacement, tortuosity, and curvature-related indices exhibited greater sensitivity to centerline extraction and endpoint definition.

These findings illustrate the technical feasibility of VB-ROI-focused segmentation for MRA-based vertebrobasilar morphometric quantification and provide a methodological foundation for future studies on VBD and posterior circulation infarction.