The project addresses key limitations of current methods, such as fragmented vessel trees and limited generalization, by focusing on three main challenges:
Annotation quality and evaluation: Defining robust guidelines and metrics to assess and improve the quality of vascular annotations, in order to strengthen model training and validation across diverse imaging conditions.
Topology-preserving segmentation: Reformulating segmentation as a recursive vessel tracking task, enabling more anatomically faithful reconstructions by explicitly following the branching structure of vascular networks.
Semi-automatic methods guided by foundation models: Investigating interactive segmentation approaches where users provide rough centerlines, used as prompts for foundation models specialized in medical imaging to improve segmentation accuracy.
All tools will be integrated into a 3D Slicer plug-in, developed in collaboration with Kitware, to produce a user-friendly vascular segmentation tool suited for clinical and translational research. Partners CREATIS, INSA Lyon
LaTIM, IMT-Atlantique