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

Figures

• 

  Figure 1. Overview of the proposed VBD-oriented VB-ROI focused segmentation-to-morphometry workflow. The method does not introduce a new network architecture; instead, it reformulates VBD-oriented analysis from generic full-image vessel segmentation into clinically targeted BA/R-VA/L-VA segmentation and morphometric quantification. The full-image nnU-Net provides coarse anatomical localization, the prediction-derived VB-ROI constrains the fine segmentation stage, and the restored full-size mask is used for segmentation evaluation, morphometric analysis, and threshold-based consistency analysis.

• 

  Figure 2. Segmentation performance overview for VBD-relevant vessels. Compared with full-image nnU-Net, both GT-based VB-ROI and prediction-derived two-stage VB-ROI increase mean Dice, IoU, and clDice while reducing HD95 and ASD. The improvement is especially pronounced for R-VA and L-VA, indicating that ROI-focused segmentation directly addresses the instability of full-image segmentation for thin vertebral arteries. GT-based VB-ROI denotes an oracle upper-bound setting, whereas the two-stage VB-ROI setting is the practical inference workflow.

• 

  Figure 3. Representative 2D overlay visualization. Each row shows the original MRA slice, ground truth, full-image nnU-Net prediction, two-stage VB-ROI prediction, and error map. BA, R-VA, and L-VA are shown in red, blue, and green, respectively. The error map highlights true positives, false positives, and false negatives. Compared with full-image segmentation, the two-stage VB-ROI workflow produces more stable vertebrobasilar vessel delineation with fewer discontinuities and boundary outliers. Source: TopBrain Challenge Data Release, Zenodo, doi: 10.5281/zenodo.16878417.

• 

  Figure 4. Three-dimensional surface rendering of BA, R-VA, and L-VA. The GT mask, full-image nnU-Net prediction, two-stage VB-ROI prediction, and error map are shown from the same viewpoint. The 3D visualization highlights the tubular continuity of the vertebrobasilar system and demonstrates that ROI-focused segmentation better preserves vessel geometry, which is critical for centerline and diameter measurements. Source: TopBrain Challenge Data Release, Zenodo, doi: 10.5281/zenodo.16878417.

• 

  Figure 5. Prediction-derived VB-ROI localization analysis. The figure shows the original MRA, GT overlay, Stage-1 coarse prediction, prediction-derived ROI bounding box, and cropped ROI view. Stage 1 is used for anatomical localization rather than final segmentation. The resulting ROI provides a focused field of view for Stage-2 fine segmentation while preserving the BA/R-VA/L-VA region required for VBD-oriented analysis. Source: TopBrain Challenge Data Release, Zenodo, doi: 10.5281/zenodo.16878417.

• 

  Figure 6. ROI margin sensitivity analysis. A small physical margin already allows the prediction-derived VB-ROI to fully enclose the target vertebrobasilar vessels, whereas larger margins progressively enlarge the crop volume without improving coverage.

• 

  Figure 7. Continuous VBD-oriented morphometric measurements derived from ground-truth and prediction masks. The figure compares GT-derived and prediction-derived values for BA diameter, BA centerline measurements, VA centerline length, and VA lateral shift. Diameter-related measurements show close agreement, supporting the use of the proposed workflow for diameter-based VBD-oriented assessment. Length and shift are computed from annotated vessel segments and should be interpreted as segmentation-derived geometric measurements rather than complete clinical anatomical measurements of the entire vertebrobasilar course.

• 

  Figure 8. Failure case analysis of the two-stage VB-ROI workflow. Representative cases are selected according to low Dice, low clDice, high HD95, or high component error. Each row shows the ground truth, two-stage prediction, error map, and local zoom. False negatives are shown in cyan and false positives in magenta. Enlarged views highlight the regions indicated by dashed boxes. Residual errors are mainly associated with distal vessel endpoints, weak VA signal, local centerline fragmentation, side-branch ambiguity, or small boundary deviations that disproportionately affect HD95.