⚡ electrical engineering

Scaling up fine-grained intracranial vessel annotations in computed tomography angiography

This paper introduces SemanticVessel, a large-scale dataset for fine-grained intracranial vessel segmentation in 4D-CTA scans that leverages multi-phase data reuse and expert annotation to create a generic arterial class, demonstrating that including minor arteries improves segmentation performance across all vascular territories.

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

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC BY 4.0

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

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 your brain's blood vessels as a massive, intricate city of roads. Some are huge highways (like the main arteries), while others are tiny, winding alleyways that reach every single neighborhood. For doctors to diagnose problems like strokes or aneurysms, they need a perfect map of this city.

Currently, creating this map is like trying to draw every single street in a giant city by hand, one brick at a time. It takes radiologists (the expert mapmakers) hours of staring at 3D scans, which is slow, expensive, and prone to human error.

This paper introduces a new project called SemanticVessel, which is like a "smart assistant" for drawing these maps, specifically for a type of scan called a CTA (Computed Tomography Angiography). Here is how they did it, explained simply:

1. The Magic of "Time-Lapse" Photography

Most brain scans are like a single snapshot. But the researchers used a special "4D" scan, which is more like a time-lapse video.

The Analogy: Imagine taking a photo of a city where the traffic lights change. First, the "artery" cars (red) drive through, and then the "vein" cars (blue) follow.
The Trick: Because the researchers have this video, they can use software to subtract the "blue" traffic from the "red" traffic and vice versa. This creates a super-clear image where the arteries stand out alone, and the veins stand out alone, without the background noise.

2. The "Grow Your Own Garden" Method

Instead of asking a human to trace every tiny vessel from scratch, the team built a system that does the heavy lifting:

Step 1: The computer finds the obvious, thick "trunks" of the trees (the main vessels).
Step 2: It uses a technique called "region growing." Think of this like planting a seed in a garden. Once the computer finds a seed (a clear vessel point), it lets the "plant" grow outward, following the path of the blood flow, until it hits a wall or runs out of "water" (contrast).
Step 3: A human expert (a radiologist) only steps in to check the garden and put name tags on the main branches. They don't have to draw the leaves; they just label the branches.

3. The "Generic" Class Problem

In the past, if a computer couldn't figure out a tiny, distant vessel, it would just throw it in the "trash" (background) category. The researchers realized this was like telling a GPS, "Ignore all the small side streets." This confused the AI.

The Fix: They created a "catch-all" category called "Other Arteries." Even if the computer didn't know the specific name of a tiny vessel, it was told to label it as "Artery" rather than "Background." This taught the AI that everything in that area is a road, even if it doesn't know the street name.

4. The Results: A Better Map

The team trained their AI using these new, smarter maps and tested it against other existing datasets.

The Outcome: The AI that was taught to recognize "generic" arteries and veins performed much better than those that ignored them. It didn't just get the big highways right; it found the tiny alleyways that other models missed.
The "Left-Right" Clue: They also gave the AI a special hint: a coordinate system that tells it "Left" is -1 and "Right" is +1. Since the brain is symmetrical (like having a left and right hand), this helped the AI understand that if it sees a big artery on the left, there's likely a matching one on the right. This made the maps even more accurate.

5. What They Found (and Didn't Find)

Success: The system successfully created detailed maps for 41 patients (360 scans) and proved that using the "time-lapse" 4D scans makes the job much faster and the results better.
Limitations: The paper admits that while the AI is great, it sometimes misses very complex areas where the brain's "roads" cross over each other in confusing ways. Also, because the maps were made by a computer with human checks, they aren't 100% perfect for simulating blood flow (which requires a perfectly connected pipe system), but they are excellent for identifying what vessels are there.

In a nutshell: The researchers built a tool that turns a slow, manual drawing job into a semi-automated process. By using time-based scans and teaching the AI to recognize "unknown" vessels as valid roads rather than background noise, they created a much larger and more accurate library of brain vessel maps than ever before. They are sharing this tool and the maps with the scientific community to help others build better brain maps in the future.

Technical Summary: Scaling up Fine-Grained Intracranial Vessel Annotations in Computed Tomography Angiography

Problem Statement
Intracranial vessel segmentation is essential for diagnosing and managing cerebrovascular diseases (CVD), yet it remains largely a manual, time-consuming process performed by radiologists. While Computed Tomography Angiography (CTA) is widely used in clinical practice due to its cost-effectiveness and speed compared to Magnetic Resonance Angiography (MRA), existing deep learning approaches for vessel segmentation have predominantly focused on MRA data. Furthermore, current CTA datasets often treat distal or minor arteries as background, limiting the model's ability to perform fine-grained semantic segmentation. The lack of large-scale, fine-grained annotated datasets for CTA hinders the development of automated tools that can provide quantitative biomarkers for disease assessment.

Methodology
The authors propose SemanticVessel, a semi-automated pipeline and dataset designed to scale fine-grained annotations for intracranial vessels in dynamic 4D-CTA scans.

Data Acquisition and Preprocessing: The study utilizes dynamic 4D-CTA data, which captures contrast dynamics across arterial and venous phases. The preprocessing pipeline involves:
1. Registration: Aligning arterial and venous phases to a common spatial frame to correct for motion.
2. Phase Separation: Applying subtraction techniques to suppress background and non-target vessels. Specifically, vein suppression is applied to the arterial phase (retaining voxels where $S_a - S_{v\to a} > 0$ ) and artery suppression to the venous phase, creating distinct volumes ( $S^*_a$ and $S^*_v$ ).
3. Seed Generation: Using Nyuz histogram normalization, Renyi entropy-based auto-thresholding, and multi-scale Hessian vesselness filters to generate initial binary vessel masks.
4. Region Growing: An intensity-guided region growing algorithm expands these seeds to recover thin vessels, with specific tolerances defined to prevent over- or under-segmentation.
Annotation Strategy:
- Expert Intervention: An experienced neuro-radiologist uses an interactive GUI to assign semantic labels to the extracted vessel centerlines.
- Class Definition: The dataset includes 20 unique arterial classes (e.g., ICA, M1, A1, VA, BA, P1, PComm) and a single "vein" class. Crucially, minor arteries that are too costly to annotate individually are merged into a generic "other artery" class rather than discarded as background.
- Label Propagation: Due to the multi-phase nature of 4D-CTA, labels generated for one phase are rigidly registered and propagated to other phases within the same series, significantly increasing the dataset size without additional annotation cost.
- Dataset Scale: The final dataset comprises 360 CTA volumes from 41 patients for training, plus 50 scans from 5 held-out patients for manual verification.
Model Training:
- The authors employ a 3D nnU-Net architecture with a Residual Encoder Large (ResEncL) preset.
- Input: The model takes two channels: the CTA image and a subject-specific left-right coordinate volume (encoding normalized distance from the brain midline) to help distinguish hemispheric homologous structures.
- Loss Functions: Models are trained using a combination of Dice and Cross-Entropy losses. Additionally, a "skeleton recall loss" is utilized to penalize missed centerline voxels, improving topological connectivity.

Key Contributions

SemanticVessel Dataset: The creation of the largest dataset to date for fine-grained brain artery segmentation in CTA, featuring 22 classes (21 arterial + 1 venous) and 360 volumes.
Novel Annotation Pipeline: A semi-automated workflow leveraging 4D-CTA dynamics to reduce the radiologist's burden, shifting the task from manual voxel annotation to label assignment on segmented centerlines.
Handling of Minor Vessels: A strategic approach to merging minor arteries into a generic class rather than background, which the authors demonstrate improves overall segmentation performance.
Cross-Phase Label Reuse: A method to multiply dataset size by reusing annotations across different temporal phases of the same scan.

Results
The models were evaluated on the TopCoW and TopBrain datasets using Modified Dice Coefficient (mDC), Average Directed Hausdorff Distance (adHD), and Topology Sensitivity (tSens).

Impact of Label Completeness: Training with "All labels" (including veins and the generic "other artery" class) yielded the best results. Specifically, the "All labels" configuration achieved the highest tSens (0.9387) and lowest adHD (0.2399) among configurations. The inclusion of the generic artery class provided the most significant supervisory benefit, improving the mDC for arteries from 0.7279 (Named arteries only) to 0.8847 (All arteries).
Loss Function: Models trained with Skeleton Recall Loss showed consistent improvements in topological completeness and boundary accuracy compared to those trained with standard Dice/CE loss.
Multi-Phase vs. Single-Phase: Training on multiphase data outperformed arterial-phase-only training, resulting in a 0.0409 improvement in mDC on the TopBrain dataset.
Left-Right Coordinates: Incorporating the left-right coordinate input significantly improved performance, particularly for paired structures like the ICA and M1, increasing the mDC for arteries from 0.6453 to 0.8734.
Generalization: Models trained on SemanticVessel outperformed models trained on subsets of TopBrain when evaluated on held-out TopBrain cases.

Significance and Claims
The paper claims that SemanticVessel addresses the critical gap in fine-grained intracranial vessel segmentation for CTA. By demonstrating that including "unlabeled" or generic vessel classes prevents the model from suppressing true vascular structures, the authors argue that their approach yields more robust and clinically relevant segmentations. The work highlights that dynamic 4D-CTA data, when processed correctly, can drastically reduce the annotation burden while increasing data volume. The authors state that their results show models trained with these additional generic classes produce better fine-grained segmentations across the board. They aim to make their code, annotation GUI, and model weights available to the scientific community to inspire the development of more comprehensive datasets and novel pre-training strategies for cerebrovascular data.

Limitations
The authors acknowledge several limitations:

The reliance on an off-the-shelf brain segmentation tool for generating the left-right coordinate channel means performance may degrade if the brain mask fails (e.g., in skull-stripped scans).
The annotation scheme is less granular than TopBrain (21 arterial structures vs. 34), which may limit utility for specific tasks like grounded vision-language modeling.
The dataset lacks coverage of occluded arteries and common anatomical variants (e.g., missing posterior communicating arteries).
The region growing step uses heuristic parameters that may require tuning for different scanners.
Annotations are not always fully connected, which limits the dataset's utility for dynamic fluid simulations requiring complete vessel trees.