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

TL;DR

The paper addresses automatic segmentation of intracranial vessels in CT angiography by leveraging dynamic 4D-CTA to subtract bone and soft tissue, enabling robust artery/vein labeling across contrast phases. It introduces DynaVessel, a dataset created via a multi-phase annotation pipeline, and trains a phase-robust nnUNet model using multi-phase GTs. The model achieves state-of-the-art performance across arteries and veins, with $mDC$ values of $0.846$ for arteries and $0.957$ for veins on TopBrain, and favorable $adHD$ and $tSens$ metrics, outperforming baselines on internal and external test sets. This approach reduces annotation burden and provides a robust vascular segmentation tool suitable for clinical workflows and downstream cerebrovascular analyses.

Abstract

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: https://github.com/alceballosa/robust-vessel-segmentation

Figures (8)

• Figure 1: Summary of our methodology. (a) Dynamic CTA are collected for all patients. (b) The subtracted images are processed to suppress voxels outside the head and resampled, then the artery-only and vein-only volumes are computed. (c) The vessel-separated images are processed using iCafe's vessel tracing algorithm and then manually validated. (d) For model training, all available CTA data (i.e., 3-19 volumes per patient) are paired with the GT segmentations.
• Figure 2: Dynamic CTA from a sample patient. The top axis shows the acquisition order of the scans. The red and blue arrows point to the arterial and venous phase images, respectively. The CTA row depicts the original CTA images, while the subtracted CTA row displays the bone/soft tissue-subtracted versions.
• Figure 3: Pre-processing steps for a dynamic CTA from a single patient. For both phases, a head ROI is used on the CTA and the subtracted CTA to suppress signals from outside the skull. Next, the subtracted images are processed using Algorithm 1 to create the vessel-separated images.
• Figure 4: Comparison of vascular structure coverage across datasets, with arteries in red and veins in blue: (a) TopCow, (b) VesselVerse, (c) TopBrain, (d) DynaVessel train, (e) DynaVessel test. Visualizations created on 3D Slicer.
• Figure 5: mDC and adHD scores across contrast phases for our model, $M_{DV}$.