💻 computer science

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

This study presents a robust deep learning methodology for 3D brain vessel segmentation in CTA scans that leverages dynamic 4D-CTA data to enhance manual annotation efficiency and expand the training dataset, resulting in a model that significantly outperforms existing benchmarks in accuracy, morphological precision, and topological sensitivity.

Original authors: Alberto Mario Ceballos-Arroyo, Shrikanth M. Yadav, Chu-Hsuan Lin, Jisoo Kim, Geoffrey S. Young, Lei Qin, Huaizu Jiang

Published 2026-02-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Alberto Mario Ceballos-Arroyo, Shrikanth M. Yadav, Chu-Hsuan Lin, Jisoo Kim, Geoffrey S. Young, Lei Qin, Huaizu Jiang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to find a specific set of tiny, winding rivers (your brain's blood vessels) inside a dense, foggy forest (your skull and brain tissue) using a series of black-and-white photos. That is essentially what doctors face when trying to map brain vessels using standard CT scans. The "fog" of bone and soft tissue often hides the rivers, making it incredibly hard to see them clearly.

This paper introduces a new, smarter way to clear that fog and teach computers to map these rivers automatically. Here is how they did it, broken down into simple steps:

1. The "Time-Lapse" Trick

Instead of taking just one snapshot of the brain, the researchers used a special scanner that takes a movie (called 4D-CTA) over about a minute.

The Analogy: Think of pouring blue dye into a garden hose system.
- At first, the dye is only in the main pipes (arteries).
- A few seconds later, it flows into the smaller branches.
- Finally, it fills the return pipes (veins).
The Innovation: Because they have the whole "movie," they can mathematically subtract the "background" (the static trees and rocks, or in this case, bone and soft tissue) by comparing the first frame (before dye) to the later frames. This leaves them with a crystal-clear view of just the flowing dye (the blood vessels).

2. The "Magic Filter" for Labeling

Once they have these clear images, they need to teach a computer what is an artery and what is a vein. Usually, this is like trying to draw a map of every single street in a city by hand—it takes forever.

The Solution: They used a clever software trick (Algorithm 1) that acts like a color-sorting machine. Since they know exactly when the dye hits the arteries vs. the veins, the software can separate the "artery movie" from the "vein movie."
The Result: They created a new dataset called DynaVessel. By using the whole "movie" instead of just one frame, they effectively multiplied their training data by 4 or 5 times without needing more patients. It's like taking one photo of a car and using it to teach a self-driving car how to drive in rain, snow, and sunshine by simulating those conditions.

3. The "Student" and the "Test"

They taught a smart computer program (a deep learning model called nnUNet) using their new, super-clear "DynaVessel" dataset. Then, they put this student to the test against other "students" (models trained on older, smaller, or less detailed datasets).

The Results:

The Old Students: When asked to find veins, the old models were almost blind, missing 99% of them. When finding arteries, they were okay but often made mistakes.
The New Student (DynaVessel): This model was a superstar.
- It found 95.7% of the veins (compared to near 0% for others).
- It found 84.6% of the arteries.
- It was incredibly precise, missing very few tiny branches and not getting confused by the background noise.

4. Why "One Photo" Isn't Enough

The researchers also ran a test to see if training on just one "perfect" moment (like only the peak arterial phase) was enough.

The Finding: If you only teach the computer using the "artery peak" photo, it gets confused when it sees a "vein peak" photo later. It's like teaching someone to drive only in the rain; they might crash when the sun comes out.
The Lesson: By training the computer on the entire movie (all the different phases), it learned to recognize the vessels no matter what the lighting (contrast) looked like. This made the model much more robust and reliable.

In Summary

The paper claims that by using a "time-lapse" approach to strip away the background noise and separating arteries from veins automatically, they created a much better training dataset. This allowed them to build a computer model that is significantly better at finding and mapping brain blood vessels than any previous method, especially for veins, which have historically been very difficult to see.

They are making their code and the model available for others to use, essentially handing the keys to a much sharper "map-making" tool for brain imaging.

Technical Summary: Robust Automatic Brain Vessel Segmentation in 3D CTA Scans using Dynamic 4D-CTA Data

Problem Statement
Intracranial computed tomographic angiography (CTA) is a critical diagnostic tool for detecting vascular abnormalities. While dynamic 4D-CTA provides detailed temporal vascular information, the predominant clinical protocol remains single-phase CTA. Robustly extracting brain vasculature from single-phase scans is challenging due to the presence of background bone and soft tissue, as well as the difficulty in distinguishing arteries from veins without extensive post-processing. Furthermore, existing deep learning approaches for vessel segmentation face significant limitations:

Data Scarcity: High-quality, fine-grained annotated datasets are rare. Existing datasets like TopCoW focus primarily on the Circle of Willis (CoW) arteries, ignoring broader vascular regions relevant to cerebrovascular disease management.
Annotation Burden: Manual annotation of CTA images requires specialized anatomical knowledge and is labor-intensive, particularly for separating arteries and veins.
Generalization: Models trained on limited modalities or single contrast phases often fail to generalize across different contrast phases or capture both arterial and venous structures accurately.

Methodology
The authors propose a novel pipeline, DynaVessel, which leverages dynamic 4D-CTA data to create a robust training dataset and improve segmentation performance. The methodology consists of three main stages:

Data Acquisition and Pre-processing:
- Dynamic CTA: Scans are acquired using a 320-detector-row CT scanner, capturing images at 19 time points over 60 seconds following contrast injection.
- Bone/Soft Tissue Subtraction: The scanner utilizes a baseline image (acquired before contrast reaches the brain) to subtract bone and soft tissue from subsequent images, enhancing vessel visualization.
- Vessel Separation: A custom algorithm (Algorithm 1) aligns the arterial and venous phase images using rigid transformations. It then suppresses venous structures in the arterial phase and arterial structures in the venous phase by comparing voxel intensities between the two phases. This results in two distinct volumes per patient: a bone- and vein-suppressed arterial volume ( $S^*_a$ ) and a bone- and artery-suppressed venous volume ( $S^*_v$ ).
- ROI Extraction: A binary mask of the skull and neck is applied to suppress extracranial tissues, and images are resampled to an isotropic resolution of 0.468 mm.
Ground Truth (GT) Annotation:
- The separated vessel volumes are processed using the iCafe feature extraction pipeline. This involves intensity normalization, thresholding (Phansalkar for tracking, Renyi entropy for visualization), and an open-curve active contour model to trace vessel-like structures.
- Manual Validation: A neuroradiologist manually validates the traces, removing spurious structures outside the skull and correcting anatomical variations. For veins, distal branches are manually refined to ensure accuracy.
- Dataset Construction: The final dataset, DynaVessel, comprises 275 head CTA scans from 43 patients. The training set includes 110 images from 29 patients, while the test set includes 165 images from 14 patients. Crucially, the training data pairs multiple contrast phases (3–19 volumes per patient) with a single set of GT annotations, effectively enlarging the dataset 4–5 times and inducing robustness to contrast variability.
Model Training and Evaluation:
- Architecture: An nnUNet (Residual Encoder UNet) model is trained on the DynaVessel dataset.
- Evaluation Metrics: Performance is assessed using Modified Dice Coefficient (mDC), Average Directed Hausdorff Distance (adHD), and Topology Sensitivity (tSens). Evaluation is conducted on binary vessel segmentation, artery-vs-vein segmentation, and semantic segmentation across 20 specific arterial and 6 venous anatomical regions.
- Baselines: The proposed model ( $M_{DV}$ ) is compared against models trained on TopCoW, VesselVerse, and the foundation model VesselFM.

Key Contributions

Annotation Pipeline: The authors document a pipeline for annotating arteries and veins in head CTA scans using dynamic 4D-CTA data, which significantly reduces annotation time by leveraging subtraction and automated tracing.
DynaVessel Dataset: The presentation of DynaVessel, a dataset containing 275 scans with fine-grained artery and vein annotations, covering a wide range of contrast phases and anatomical regions (including distal arteries and venous sinuses).
Performance Benchmarking: The evaluation of a model trained on DynaVessel against state-of-the-art baselines, demonstrating superior segmentation performance across both arteries and veins.

Results
The model trained on DynaVessel ( $M_{DV}$ ) achieved significantly better performance than models trained on TopCoW, VesselVerse, or VesselFM across all vascular regions and datasets:

TopBrain Dataset: $M_{DV}$ achieved an average mDC of 0.846 for arteries and 0.957 for veins. In comparison, the best baseline (VesselVerse) achieved 0.455 (arteries) and 0.031 (veins).
Error Margins: The model demonstrated low error margins with an adHD of 0.304 mm for arteries and 0.078 mm for veins, and high topology sensitivity (tSens) of 0.877 (arteries) and 0.974 (veins).
Robustness: While performance naturally varied by contrast phase (e.g., lower performance in pre-contrast phases), training on multi-phase data significantly improved generalization compared to models trained on single peak phases. Ablation studies confirmed that training on a single phase severely reduced the model's ability to generalize to unseen phases.
Qualitative Analysis: Visualizations showed that $M_{DV}$ achieved near-complete coverage of arterial and venous structures with very few false negatives, whereas baseline models failed to detect most veins or produced numerous spurious segmentations.

Significance
The paper claims that this work constitutes a significant step toward the automatic segmentation of all brain vascular structures in CTA scans. By utilizing dynamic 4D-CTA data, the authors overcome the longstanding difficulties of distinguishing intracranial arteries and veins and reducing the labor-intensive nature of annotation. The resulting dataset and model offer broad applicability as a primary or auxiliary input for assessing cerebrovascular pathologies, such as occlusion localization and aneurysm detection. The authors note that while the current fine-grained semantic evaluation serves as an upper-bound approximation, the methodology provides a foundation for future work to expand annotations and develop architectures for complex semantic segmentation tasks.