SeqSeg: Learning Local Segments for Automatic Vascular Model Construction

TL;DR

SeqSeg presents a seed-initiated, sequential segmentation approach to automatic vascular model construction that operates on local subvolumes with a 3D U-Net. By iteratively tracing along vessel centerlines and handling bifurcations, SeqSeg builds a globally connected vascular segmentation and surface mesh, outperforming state-of-the-art global nnU-Net baselines in covering distal branches and maintaining topology. The method demonstrates robust generalization to unseen vascular regions and de novo data, with favorable Dice, Hausdorff, and centerline overlap metrics and improved connectivity for physics-based simulations. It leverages adaptive subvolume sizing, probabilistic fusion across overlaps, and explicit handling of vessels’ geometry to deliver simulation-ready vascular models more efficiently than global pixel-wise approaches. The work is supported by public datasets and open-source code, enabling reuse for broader vascular anatomies and future integration with multi-physics workflows.

Abstract

Computational modeling of cardiovascular function has become a critical part of diagnosing, treating and understanding cardiovascular disease. Most strategies involve constructing anatomically accurate computer models of cardiovascular structures, which is a multistep, time-consuming process. To improve the model generation process, we herein present SeqSeg (sequential segmentation): a novel deep learning based automatic tracing and segmentation algorithm for constructing image-based vascular models. SeqSeg leverages local U-Net-based inference to sequentially segment vascular structures from medical image volumes. We tested SeqSeg on CT and MR images of aortic and aortofemoral models and compared the predictions to those of benchmark 2D and 3D global nnU-Net models, which have previously shown excellent accuracy for medical image segmentation. We demonstrate that SeqSeg is able to segment more complete vasculature and is able to generalize to vascular structures not annotated in the training data.

Figures (12)

• Figure 1: A typical vascular model construction workflow involves (a) creating vessel paths by manual selection of point (b) sequential segmentation of the vessel lumen boundary at discrete cross-sections along the paths and (c) lofting these segmentation rings into a unified model. This process is described in more detail in Updegrove2017SimVascular:Simulation.
• Figure 2: When viewed locally, vasculature of different sizes and anatomical regions exhibit substantial geometric similarity. A) the pulmonary artery ($r=1.5mm$), b) the brachiocephalic artery ($r=9mm$), c) the coronary artery ($r=1mm$), d) the cerebral artery ($r=2mm$) and e) the femoral artery ($r=3mm$) are presented
• Figure 3: Overview of the tracing and segmentation algorithm with inputs of the global raw image and seed points for initialization. The algorithm takes steps, stores bifurcations in the queue during tracing, and outputs a global segmentation map for post processing
• Figure 4: Preprocessing involves extracting subvolumes along ground truth centerlines and data augmentation prior to neural network training. Thousands of samples are acquired from only a few dozen models. The neural network consists of an encoder $\mathcal{E}$ followed by a decoder $\mathcal{D}$, which outputs the predicted segmentation map used to compute loss, L, during training
• Figure 5: Automatic tracing using local surface mesh predictions for 3 steps, involving 12 calculation time steps. Centerlines are extracted and the next points are chosen to move to. These steps are subsequently assembled together to form the global vasculature model