Signed Distance Field Regression from MR Vessel Wall Imaging for Intracranial Vessel Segmentation
Abstract
Purpose
To evaluate whether signed distance field regression improves geometric fidelity in intracranial vessel segmentation from non-contrast vessel wall imaging (VWI).
Methods
We use a retrospective dataset of paired MR VWI and TOF-MRA images from 61 patients. A pretrained MRA segmentation network is used to generate reference vessel labels and Euclidian distance transform was applied to obtain reference signed distance field (SDF) maps. We propose a regression UNet model trained using a distance-weighted MAE loss that preferentially penalizes regression inaccuracies close to vessel boundaries and using eikonal regularization to enforce SDF geometric properties. The final vessel label prediction is taken as the zero sublevel set of the output signed distance field. We compare our model to baseline nnUNet segmentation models. We evaluate model performance in terms of centerline Dice, Dice, average centerline distance, average symmetric surface distance.
Results
The proposed SDF regression model achieved centerline Dice (clDice) of 0.66 ± 0.05 and an average centerline distance of 2.48 ± 0.63 mm, comparable to the nnUNet baseline (0.66 ± 0.04 and 2.59 ± 0.52 mm, respectively). A nnUNet variant incorporating skeleton recall achieved higher centerline fidelity (clDice 0.70 ± 0.05) and lower distance errors. Volumetric overlap was similar across methods, with both the proposed model and nnUNet achieving Dice scores of 0.68 ± 0.05. Average symmetric surface distances were likewise comparable, reporting 1.40 ± 0.47 mm for the proposed method versus 1.48 ± 0.44 mm for nnUNet.
Conclusion
SDF-based regression for intracranial vessel segmentation from non-contrast VWI produces performance comparable to strong nnUNet baselines, but topology-aware objectives such as skeleton recall remain more effective for improving centerline fidelity. Our findings suggest that while SDF regression provides a principled geometric alternative to voxel-wise classification and encourages spatial continuity, it is not by itself sufficient to fully resolve topological discontinuities in low-contrast VWI.