💻 computer science

Intracranial Aneurysm Classification and Segmentation via Tri-Axial ROI and Multi-Task Learning

This paper presents a multi-task learning framework combining 2D tri-axial ROI extraction with a 3D nnU-Net backbone to simultaneously classify, segment, and localize intracranial aneurysms across 13 anatomical sites and four imaging modalities, achieving 2nd place in the RSNA 2025 challenge.

Original authors: Pengcheng Shi, Kaiyuan Yang, Houjing Huang, Jiawei Chen, Yan Lu, Jiaqi Liu, Murong Xu, Bjoern Menze, Xinglin Zhang

Published 2026-06-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pengcheng Shi, Kaiyuan Yang, Houjing Huang, Jiawei Chen, Yan Lu, Jiaqi Liu, Murong Xu, Bjoern Menze, Xinglin Zhang

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 are like a complex network of highways. Sometimes, a weak spot on a highway wall balloons out, forming a dangerous bubble called an aneurysm. If this bubble pops, it's a medical emergency. Doctors need to know two things to decide how to fix it: where the bubble is exactly, and what shape it has.

Until now, computer programs could mostly just say, "Yes, there's a bubble here," or "No, there isn't." They couldn't tell you which specific highway the bubble was on, or draw a detailed map of the bubble itself versus the surrounding road.

This paper introduces a new computer system that acts like a super-smart, multi-tasking detective. Here is how it works, broken down into simple steps:

1. The "Quick Look" Strategy (Stage 1)

Imagine you have a giant, 3D block of cheese (the brain scan) and you need to find a tiny crumb inside it. The old way was to slice the whole block into thin pieces and look at every single one. That takes forever.

The authors' new method is like taking a quick peek at just three specific slices: one from the top, one from the side, and one from the front.

The Trick: They use a 2D camera to look at these three slices.
The Result: By combining what it sees from these three angles, the computer can instantly draw a 3D box around the area where the bubble might be.
The Benefit: This is 12.7 times faster than looking at the whole block slice-by-slice. It zooms in on the "neighborhood" without wasting time on the empty parts of the brain.

2. The "Two-Headed" Brain (Stage 2)

Once the computer zooms in on that specific neighborhood, it uses a powerful 3D brain (based on a famous AI model called nnU-Net) to do three jobs at once:

Classify: Is there a bubble?
Locate: Which of the 13 specific highway sections is it on? (e.g., "Left side, front artery").
Map: Draw a precise outline of the bubble and the highway it's attached to.

The "Twin Decoder" Problem:
There's a big problem in this job: The blood vessels are huge, but the aneurysm bubble is tiny. If you ask a computer to learn both at the same time, it gets lazy and ignores the tiny bubble because the big vessels are so much easier to see.

The Solution: The authors built a "twin decoder" system.

Decoder A focuses on finding the big vessels and a "blob" that represents any bubble.
Decoder B focuses on matching specific vessels to their specific bubble locations.
The Magic: By splitting the work, the computer doesn't get overwhelmed. It treats the tiny bubble as a VIP guest, giving it special attention so it doesn't get lost in the crowd of big vessels.

3. Handling Different "Cameras"

The system has to work with four different types of medical scans (CTA, MRA, T2, T1-post). Think of these like four different camera lenses that take pictures of the same scene but with different lighting and colors.

The system includes a special "translator" that learns to recognize which camera was used and adjusts its understanding accordingly, so it works perfectly no matter which scan type it sees.

4. The Results

Speed: Because of the "Quick Look" strategy, the system is incredibly fast.
Accuracy: In a major competition (RSNA 2025), this system came in 2nd place.
Detail: It is the first system to successfully do all three things at once: say "Yes/No," name the exact location, and draw the detailed map of both the vessel and the bubble.

What They Did Not Claim

The authors are very honest about what they didn't do:

They didn't test this on real patients in a hospital to see if it saves lives (yet).
They didn't create perfect "ground truth" maps for the bubbles. Since the competition data only gave them the center point of the bubble (like a pin on a map), they had to use a smart guessing game (called "active learning") to draw the full shape of the bubble. They manually checked and corrected these guesses, but they admit the bubble shapes are still an estimate, not a perfect medical fact.

The Bottom Line

The authors built a fast, smart tool that can zoom in on brain scans, find tiny aneurysms, tell you exactly where they are on the brain's "highway map," and draw them out, all while handling different types of medical images. They made the code and the tool available for others to use and improve.

Technical Summary: Intracranial Aneurysm Classification and Segmentation via Tri-Axial ROI and Multi-Task Learning

Problem Statement
Intracranial aneurysms are often asymptomatic until rupture, a event carrying high mortality. Effective risk assessment and treatment planning require precise knowledge of both aneurysm morphology and anatomical location. However, existing automated methods are limited to binary outputs (presence/absence, binary masks) and fail to provide fine-grained anatomical classification or multi-class segmentation. Furthermore, the heterogeneity of imaging modalities (CTA, MRA, T2, T1-post) presents challenges regarding contrast mechanisms, voxel spacings, and fields of view. Current Region of Interest (ROI) extraction methods often rely on modality-specific preprocessing or computationally expensive 3D sliding-window inference, limiting scalability.

Methodology
The authors propose a two-stage framework designed to perform multi-label classification, multi-class aneurysm segmentation, and multi-class vessel segmentation across 13 anatomical locations and four imaging modalities.

Stage 1: 2D Tri-Axial ROI Extraction
To avoid the computational cost of full 3D inference, the authors introduce a fast 2D tri-axial approach. For an input volume, the system extracts three slices per orthogonal axis (axial, sagittal, coronal) at positions {1/4, 1/2, 3/4}, totaling up to nine slices. These slices are processed by a standard 2D nnU-Net binary segmentation model. The resulting 2D masks constrain spatial dimensions (e.g., axial slices constrain $y$ and $x$ ). The 3D bounding box is derived by averaging the coordinates of the non-empty predictions across the axes. This method reduces inference to nine slice predictions per volume, achieving a 12.7× speedup compared to 3D sliding-window inference while consistently encompassing the Circle of Willis and proximal arterial segments.
Stage 2: 3D Multi-Task Learning with Dual-Decoders
The cropped ROI volumes are resized to $224 \times 224 \times 224$ and processed by a 3D multi-task network based on nnU-Net.
- Dual-Decoder Design: To address the extreme volume imbalance where aneurysm voxels are 2–3 orders of magnitude fewer than vessel voxels, the network employs two decoders rather than a single 26-class decoder.
  - Decoder 1: Outputs 14 channels, separating 13 vascular anatomy classes from a single merged aneurysm class.
  - Decoder 2: Outputs 13 channels, pairing each vessel anatomy class with its corresponding aneurysm location class.
  - Reconstruction: Post-training, the full 26-class segmentation is recovered by fusing outputs: if Decoder 1 predicts the merged aneurysm class and Decoder 2 predicts vessel anatomy class $k$ , the final label is the aneurysm location class $k+13$ .
- Classification Heads: Three classification heads utilize cross-attention pooling on the encoder bottleneck to predict: (1) input modality (4 classes), (2) aneurysm presence (binary), and (3) multi-label aneurysm location (13 classes).
- Loss Function: The total loss is a weighted sum of segmentation losses (Dice + weighted Cross-Entropy) and classification losses (BCE/CE). Aneurysm-associated classes receive significantly higher weights (10× in Decoder 1, 5× in Decoder 2) to counter class imbalance.
Data Preparation and Augmentation
Since the RSNA-2025 challenge provided only center-point coordinates for aneurysms and not voxel-level masks, the authors employed an active-learning strategy. They generated pseudo-labels for vessel anatomy and aneurysm masks using initial models, which were manually reviewed, corrected, and fed back iteratively. This process covered all 4,348 training series. Additionally, left-right label-swap augmentation was applied to double the dataset size for cases with lateralized labels.

Key Contributions

2D Tri-Axial ROI Extraction: A strategy that handles multi-modal neurovascular data without full 3D segmentation, reducing inference to nine slice predictions per volume.
Unified Multi-Task Framework: The first method to unify multi-label aneurysm classification, multi-class aneurysm segmentation, and multi-class vessel segmentation at multi-class granularity across four modalities.
Dual-Decoder Architecture: A design specifically addressing the extreme volume imbalance between aneurysm and vessel classes, allowing for effective gradient flow for sparse aneurysm targets.
Public Release: The authors released code, model weights, and a 3D Slicer plugin to support further research.

Results
The framework achieved 2nd place in the RSNA 2025 Intracranial Aneurysm Detection challenge.

Performance: The final ensemble (two folds) achieved a Public AUC of 0.90035 and a Private AUC of 0.86727.
Ablation Insights:
- Refining segmentation annotations and expanding vessel anatomy labels via active learning yielded a +0.032 gain on the private leaderboard.
- Test-Time Augmentation (TTA) with 8× combinations provided a +0.010 gain.
- Ensembling two folds provided a marginal but consistent +0.005 gain.
Efficiency: The 2D tri-axial approach achieved an average inference speed of 0.51 seconds per volume on an RTX 3090, compared to 6.44 seconds for a 3D sliding-window baseline.

Significance and Claims
The paper claims that its primary significance lies in moving beyond binary detection to provide fine-grained, multi-class anatomical outputs across heterogeneous imaging modalities. The authors highlight that their dual-decoder design offers a viable strategy for handling extreme class imbalance in medical segmentation tasks where target structures (aneurysms) are minute relative to the background (vessels). They note that while the segmentation evaluation remains qualitative due to the lack of ground-truth voxel masks in the challenge, the framework successfully produces dense anatomical outputs even when trained primarily for classification. The authors modestly acknowledge limitations, including the reliance on pseudo-labels generated via active learning and the non-end-to-end nature of the 26-class segmentation recovery, which requires post-processing fusion of two decoders.