Cascaded multitask U-Net using topological loss for vessel segmentation and centerline extraction
Pierre Rougé, Nicolas Passat, Odyssée Merveille
TL;DR
This work tackles topology-preserving vessel segmentation and centerline extraction in 3D angiography by introducing a cascaded multitask U-Net that learns both vessel segmentation and skeletonization, enabling topological constraints via the clDice loss. A learned skeletonization network replaces suboptimal soft-skeleton approaches, with the final loss combining Dice on segmentation, Dice on skeletons, and clDice topology guidance, defined as $clDice = 2 \cdot \frac{T_{prec}(C_P,S_G) \cdot T_{sens}(C_G,S_P)}{T_{prec}(C_P,S_G) + T_{sens}(C_G,S_P)}$. Evaluations on 34 brain MRA volumes with 5-fold cross-validation show that the cascaded U-Net achieves comparable Dice, improved clDice and topological metrics, and competitive runtimes relative to skeletonization baselines, confirming that joint segmentation-skeletonization with topology-aware loss yields more accurate vascular topology. Hyperparameter exploration favors equal weighting of segmentation and skeletonization losses $(\lambda_1, \lambda_2) = (0.5,0.5)$ with a frozen skeletonization network during joint training, indicating robust topology improvement without excessive computation. The approach advances clinically relevant vessel analysis by delivering topology-correct vessel masks and centerlines suitable for downstream hemodynamic modeling, while maintaining practical runtimes.
Abstract
Vessel segmentation and centerline extraction are two crucial preliminary tasks for many computer-aided diagnosis tools dealing with vascular diseases. Recently, deep-learning based methods have been widely applied to these tasks. However, classic deep-learning approaches struggle to capture the complex geometry and specific topology of vascular networks, which is of the utmost importance in most applications. To overcome these limitations, the clDice loss, a topological loss that focuses on the vessel centerlines, has been recently proposed. This loss requires computing, with a proposed soft-skeleton algorithm, the skeletons of both the ground truth and the predicted segmentation. However, the soft-skeleton algorithm provides suboptimal results on 3D images, which makes the clDice hardly suitable on 3D images. In this paper, we propose to replace the soft-skeleton algorithm by a U-Net which computes the vascular skeleton directly from the segmentation. We show that our method provides more accurate skeletons than the soft-skeleton algorithm. We then build upon this network a cascaded U-Net trained with the clDice loss to embed topological constraints during the segmentation. The resulting model is able to predict both the vessel segmentation and centerlines with a more accurate topology.