203. Mind the Gap: Mesh-Guided Repair of Broken Vessels

Abstract:

Vessel segmentation is usually trained and evaluated voxel by voxel, yet a missing bridge of a few voxels can split a vessel into pieces while barely changing Dice, breaking downstream centreline extraction, branch labelling and flow simulation. We present Mind the Gap, a post-processing framework that repairs the connectivity of an existing nnU-Net vessel mask without retraining and without image intensities or probability maps. For each predicted mask, an anatomy-specific template mesh (tube, torus or sphere) is fitted to the mask surface by a four-stage graph-convolutional decoder, optimised per case with a Chamfer loss and geometric regularisers. The mesh is never voxelised as the output: it only proposes or validates thin bridges between disconnected components, each accepted only if it merges components within a strict foreground-growth limit. On out-of-fold nnU-Net predictions for four public datasets covering three anatomies, SEGA and AortaSeg24 (aorta, n = 145), TopCoW (Circle of Willis, n = 125) and PARSE (pulmonary arteries, n = 100), Dice changed by at most 0.003, whereas connected-component Dice rose from 0.60 to 0.99, 0.72 to 0.84 and 0.03 to 0.86, respectively, and the mean number of components fell to 1.0–1.6. The main cost was in the Circle of Willis, where the false-branch fraction rose from 0.040 to 0.063 and circles disconnected in the reference were usually connected. A first-order meta-learned (FOMAML) initialisation reaches the fit quality of 500 standard steps within 178–280 steps, bringing fitting and repair to about 30 s per aortic case. Explicit meshes can thus provide a practical geometric prior for repairing vascular topology.

Gniewosz Drwięga¹, Wojciech Szymański^1,2, Marek Wodziński^1,2

¹ Sano – Centre for Computational Personalised Medicine International Research Foundation, Kraków, Poland

² AGH University of Krakow, Poland

ShapeMI 2026 @ MICCAI · Best Poster Award

Keywords: vessel segmentation · topology · mesh deformation · geometric deep learning · nnU-Net

About the author:

Gniewosz Drwięga holds a PhD in Systems Neurobiology from Jagiellonian University, where he also worked as an Assistant Professor in the Department of Computational Biophysics and Bioinformatics. His scientific interests include computer vision, medical image analysis, and deep learning.

Preprint: arXiv:2609.29779 · Code: github.com/SanoScience/mind-the-gap