Towards the Automatic Segmentation, Modeling and Meshing of the Aortic Vessel Tree from Multicenter Acquisitions: An Overview of the SEG.A. 2023 Segmentation of the Aorta Challenge

TL;DR

The SEG.A. 2023 MICCAI satellite challenge tackles automated segmentation of the aortic vessel tree (AVT) from multicenter CTA data by providing a large, publicly available training set and a private test set to benchmark state-of-the-art methods. The study reveals a convergence toward 3D U-Net–based architectures, with ensemble approaches delivering the strongest overall performance, and shows that careful preprocessing, data augmentation, and post-processing substantially influence results. A comprehensive evaluation framework combines segmentation metrics ($DSC$, $HD$) with global sensitivity analyses via Sobol' indices to assess robustness across image variations, while surface meshing tasks demonstrate practical pipelines for visualization and clinical-grade modeling. The paper delivers a benchmark and open resources to drive future robust, clinically translatable AVT analysis tools, highlighting both achieved progress and remaining challenges in generalization, small-branch detection, and topology preservation.

Abstract

The automated analysis of the aortic vessel tree (AVT) from computed tomography angiography (CTA) holds immense clinical potential, but its development has been impeded by a lack of shared, high-quality data. We launched the SEG.A. challenge to catalyze progress in this field by introducing a large, publicly available, multi-institutional dataset for AVT segmentation. The challenge benchmarked automated algorithms on a hidden test set, with subsequent optional tasks in surface meshing for computational simulations. Our findings reveal a clear convergence on deep learning methodologies, with 3D U-Net architectures dominating the top submissions. A key result was that an ensemble of the highest-ranking algorithms significantly outperformed individual models, highlighting the benefits of model fusion. Performance was strongly linked to algorithmic design, particularly the use of customized post-processing steps, and the characteristics of the training data. This initiative not only establishes a new performance benchmark but also provides a lasting resource to drive future innovation toward robust, clinically translatable tools.

Figures (7)

• Figure 1: Overview of the challenge phases and the geographic distribution of participants. The SEG.A.challenge was structured into three main phases: Phase 1 – participants tested their Docker images and algorithms using two hidden cases; Phase 2 – five hidden cases were used for evaluation; Phase 3 – the top three teams from Phase 2 were given the opportunity to resubmit their final algorithms, with the ultimate winner receiving a monetary prize. Additionally, two optional subtasks, surface meshing and refinement, were encouraged. The challenge attracted 733 registrations from all continents. In the first phase, 17 teams submitted their algorithms, while in the second phase, 14 teams participated.
• Figure 2: Technical overview of leaderboard submissions. All participants adopted deep learning-based approaches to address the challenge task. Most approaches were built upon three-dimensional U-Net architecture, with the Dice loss function serving as the primarily optimization objective. CNN = convolutional neural network, VIT = vision transformer, CBAM = convolutional block attention module, DSC = dice similarity coefficient, CE = cross-entropy, FL = Focal loss, HDL = Hausdorff distance loss. Top: Model architecture backbones. Bottom: Loss functions.
• Figure 3: An overview of algorithm performance is presented, with challenge outcomes evaluated using DSC and HD. Results are ordered from best (left) to worst (right). Each dot represents the mean value of the metric for each of the three challenge phases.
• Figure 4: Qualitative examples of automated AVT segmentation. From left to right: the ground truth mask, the segmentation produced by Attention U-Net, the segmentation produced by nnU-Net, and the segmentation produced by M3F byeon2023m3f.
• Figure 5: Final rankings of the two subtasks.