Geometry-Preserving Intracranial Vessel Tree Extraction from MR Vessel Wall Imaging
Abstract
Purpose
To determine whether appearance-enhancing synthesis pipelines are sufficient for recovering a topologically coherent intracranial vessel tree from non-contrast MR vessel wall imaging (VWI), or whether explicit geometry-aware segmentation is required.
Methods
We use a retrospective dataset of paired VWI and TOF-MRA images from 61 patients, with reference vessel labels generated from TOF-MRA. We investigate three modeling paradigms for vessel tree extraction from VWI: (1) direct segmentation, (2) contrast-enhanced image synthesis followed by segmentation, and (3) hybrid synthesis approaches incorporating segmentation intent. Direct segmentation is performed using nnUNet, with and without skeleton recall loss to promote centerline awareness. In the synthesis approach, Pix2Pix GANs are trained to generate high-contrast synthetic MRA from VWI. The hybrid approach incorporates segmentation intent via total variation regularization to encourage geometric continuity and by initializing the GAN generator with a pretrained nnUNet model. Synthetic MRAs are subsequently segmented using a pretrained MRA segmentation model. All models and approaches are evaluated and compared using Dice and centerline Dice (clDice).
Results
Synthetic MRAs showed improved vessel contrast with nnUNet initialization. The incorporation of total variation regularization further enhanced vessel homogeneity and continuity. The best hybrid Pix2Pix model achieved a mean Dice and clDice of 0.55±0.03 and 0.56±0.03, respectively. The direct segmentation nnUNet model outperformed all Pix2Pix models, achieving a mean Dice score and clDice score of 0.64±0.03 and 0.62±0.03, respectively. Adding skeleton recall loss did not significantly alter Dice, but increased clDice score to 0.67±0.03.
Conclusion
Although synthesis-based approaches can enhance vessel appearance, they struggle to recover the global geometry required for coherent vessel tree extraction. Direct segmentation from VWI, especially with addition of geometry-aware loss functions, is better suited for this task. These findings underscore the importance of geometry-focused modeling and evaluation metrics when extracting intact vessel trees from VWI where vessel signal is inherently sparse.