Robust automatic brain vessel segmentation in 3D CTA scans using dynamic 4D-CTA data

In this study, we develop a novel methodology for annotating the brain vasculature using dynamic 4D-CTA head scans. By using multiple time points from dynamic CTA acquisitions, we subtract bone and soft tissue to enhance the visualization of arteries and veins, reducing the effort required to obtain manual annotations of brain vessels. We then train deep learning models on our ground truth annotations by using the same segmentation for multiple phases from the dynamic 4D-CTA collection, effectively enlarging our dataset by 4 to 5 times and inducing robustness to contrast phases. In total, our dataset comprises 110 training images from 25 patients and 165 test images from 14 patients. In comparison with two similarly-sized datasets for CTA-based brain vessel segmentation, a nnUNet model trained on our dataset can achieve significantly better segmentations across all vascular regions, with an average mDC of 0.846 for arteries and 0.957 for veins in the TopBrain dataset. Furthermore, metrics such as average directed Hausdorff distance (adHD) and topology sensitivity (tSens) reflected similar trends: using our dataset resulted in low error margins (adHD of 0.304 mm for arteries and 0.078 for veins) and high sensitivity (tSens of 0.877 for arteries and 0.974 for veins), indicating excellent accuracy in capturing vessel morphology. Our code and model weights are available online at https://github.com/alceballosa/robust-vessel-segmentation

💡 Research Summary

This paper introduces a novel pipeline for creating high‑quality brain vessel annotations by leveraging dynamic 4D‑CT angiography (CTA) data and demonstrates that models trained on these annotations substantially outperform those trained on existing CTA datasets. Conventional CTA‑based vessel segmentation suffers from two major challenges: (1) the need to manually remove bone and soft tissue, which is time‑consuming and error‑prone, and (2) limited robustness to contrast‑phase variations because most public datasets contain only a single phase per patient. The authors address both issues by exploiting the temporal dimension of 4D‑CTA. Each scan includes a baseline (pre‑contrast) volume and multiple post‑contrast volumes captured at different time points. By subtracting the baseline from each contrast phase, bone and soft tissue are effectively eliminated, leaving high‑contrast arterial and venous images. This subtraction is performed automatically using rigid registration and a custom algorithm that suppresses arterial signal in venous phases and vice‑versa, producing two clean volumes per phase: an artery‑only and a vein‑only image.

The key innovation is to reuse the same ground‑truth (GT) vessel masks across all available phases for a given patient. Because a patient typically contributes 3–19 phases, the effective size of the training set is multiplied by a factor of four to five, while simultaneously exposing the network to a wide range of contrast intensities. The resulting dataset, named DynaVessel, comprises 110 training volumes from 29 patients and 165 test volumes from 14 patients, each annotated with separate artery and vein labels. Annotations are generated by running the iCafe vessel‑tracing pipeline on the artery‑only and vein‑only volumes, followed by expert review and manual correction in 3D Slicer. For the test set, 21 arterial sub‑structures and the venous system were manually verified.

For model training, the authors employ nnU‑Net v2 with a Residual Encoder UNet (large) configuration. All volumes are rigidly aligned to the arterial phase, resampled to isotropic 0.468 mm voxels, and normalized. Training runs for 1000 epochs with a linearly decaying learning rate starting at 0.01, using the automatically selected hyper‑parameters of nnU‑Net. The network outputs a two‑class segmentation (artery, vein).

Evaluation uses three complementary metrics: (i) modified Dice coefficient (mDC), which emphasizes sensitivity; (ii) average directed Hausdorff distance (adHD), a boundary‑error metric that penalizes over‑segmentation; and (iii) topology sensitivity (tSens), which measures how much of the ground‑truth centerline is recovered. Compared with models trained on three publicly available CTA datasets of similar size—TopCoW, TopBrain, and VesselVErse—the DynaVessel‑trained nnU‑Net achieves markedly higher scores across all metrics. On the TopBrain test set, the model reaches mDC = 0.846 for arteries and 0.957 for veins, adHD = 0.304 mm (arteries) and 0.078 mm (veins), and tSens = 0.877 (arteries) and 0.974 (veins). These results represent a substantial improvement over the best competing model (nnU‑Net trained on TopBrain) which attains mDC ≈ 0.789, adHD ≈ 0.274 mm, and tSens ≈ 0.930 for arteries.

The authors also evaluate cross‑dataset generalization on the ISLES benchmark, where the DynaVessel model again outperforms models trained on TopCoW and VesselVErse, achieving mDC = 0.960 (arteries) and 0.989 (veins) with low adHD values. This demonstrates that the multi‑phase augmentation confers robustness not only within the same acquisition protocol but also across different scanners and patient populations.

In addition to performance gains, the paper contributes the DynaVessel dataset (publicly released on GitHub) and the source code for the subtraction and artery/vein suppression pipeline, facilitating reproducibility and future research. Limitations include the reliance on 4D‑CTA hardware, which is not universally available, and the need for expert validation of the automatically generated masks. Future work may explore multimodal fusion with MR angiography, automated phase selection, and lightweight models for real‑time clinical deployment. Overall, the study provides a compelling solution to the data scarcity and phase‑variability problems that have hampered CTA‑based brain vessel segmentation, and it sets a new benchmark for both arterial and venous segmentation performance.