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

Bjoern Menze; Houjing Huang; Jiaqi Liu; Jiawei Chen; Kaiyuan Yang; Murong Xu; Pengcheng Shi; Xinglin Zhang; Yan Lu

arxiv: 2606.26706 · v1 · pith:7FVCSWAHnew · submitted 2026-06-25 · 💻 cs.CV

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

Pengcheng Shi , Kaiyuan Yang , Houjing Huang , Jiawei Chen , Yan Lu , Jiaqi Liu , Murong Xu , Bjoern Menze

show 1 more author

Xinglin Zhang

This is my paper

Pith reviewed 2026-06-26 05:42 UTC · model grok-4.3

classification 💻 cs.CV

keywords intracranial aneurysmmulti-task learningROI extractionimage segmentationclassificationmedical imagingCTAMRA

0 comments

The pith

Tri-axial ROI extraction with a dual-decoder multi-task network enables simultaneous location-specific classification and segmentation of intracranial aneurysms and vessels across modalities.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a two-stage system that first uses a fast 2D tri-axial method to locate candidate regions and then runs a 3D multi-task network to classify aneurysms according to one of 13 anatomical sites while segmenting both the aneurysms and the vessels. Existing tools stop at simple detection, but this setup supplies the morphology and location details needed for rupture risk assessment and treatment planning. A dual-decoder structure and cross-attention pooling address the severe size difference between aneurysm and vessel classes and the variation across CTA, MRA, T2, and T1-post scans. The resulting two-fold ensemble reached second place in the RSNA 2025 Intracranial Aneurysm Detection challenge.

Core claim

The framework simultaneously performs multi-label classification, multi-class aneurysm segmentation, and multi-class vessel segmentation across 13 anatomical locations and four imaging modalities by combining fast 2D tri-axial ROI extraction with a 3D multi-task nnU-Net backbone whose dual-decoder design, cross-attention pooling, and modality-specific auxiliary heads mitigate volume imbalance and improve feature learning from heterogeneous inputs.

What carries the argument

Tri-axial ROI extraction paired with a dual-decoder 3D multi-task network that uses cross-attention pooling to balance aneurysm and vessel segmentation.

If this is right

The model outputs both location labels for 13 sites and separate masks for aneurysms and vessels in one forward pass.
The dual-decoder structure reduces the effect of extreme class imbalance between small aneurysms and large vessels.
Modality-specific auxiliary heads allow the same backbone to process CTA, MRA, T2, and T1-post scans without retraining.
A two-fold ensemble of the trained models placed second in the RSNA 2025 Intracranial Aneurysm Detection challenge.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Public release of the code, weights, and 3D Slicer plugin lowers the barrier for clinical groups to test the pipeline on their own scans.
Location-aware outputs could feed directly into existing rupture-risk calculators that already weight certain anatomical sites more heavily.
The multi-modality design suggests a route to training on mixed or incomplete imaging datasets for other focal vascular lesions.

Load-bearing premise

The performance improvements result from the tri-axial ROI extraction and dual-decoder architecture rather than from the two-fold ensemble or tuning choices on the challenge test set.

What would settle it

An ablation experiment on an independent held-out dataset that removes the tri-axial ROI step or the dual-decoder while keeping the ensemble shows no measurable drop in classification or segmentation scores.

Figures

Figures reproduced from arXiv: 2606.26706 by Bjoern Menze, Houjing Huang, Jiaqi Liu, Jiawei Chen, Kaiyuan Yang, Murong Xu, Pengcheng Shi, Xinglin Zhang, Yan Lu.

**Figure 1.** Figure 1: Overview of the Stage 1 tri-axial ROI extraction pipeline. A 2D nnU-Net processes slices sampled from three orthogonal axes; bounding box coordinates are extracted from each 2D prediction and averaged across axes to obtain the final 3D ROI. 2 Method [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Stage 2 multi-task framework. Dual segmentation decoders produce vessel and aneurysm location masks; three classification heads predict modality, aneurysm presence, and location. Multi-Task Loss. The total loss is a weighted sum over all five heads: L = 1 Z λsegX 2 d=1 Lsegd + L presence BCE + L location BCE + L modality CE (2) Each Lsegd combines Dice and weighted CE loss [4], with λseg = 0.5, Z = 3.… view at source ↗

**Figure 3.** Figure 3: Left-right label-swap augmentation. Same case shown in original (top) and after horizontal flipping with swapped left-right labels (bottom), each in four views (axial, coronal, sagittal, 3D mask). Aneurysm regions are highlighted with yellow dashed bounding boxes. At inference, the cropped volume is resized to 224 × 224 × 224, matching the patch size, and processed in a single forward pass. Decoders and … view at source ↗

**Figure 4.** Figure 4: Examples of the generated 26-class segmentation labels across modalities and anatomical locations. Each column depicts the same aneurysm in four views: axial slice with overlaid segmentation mask, coronal view, sagittal view, and 3D mask. Aneurysm regions are highlighted with yellow dashed bounding boxes. level anatomical labels from TopCoW [12] (with LargeIA subset excluded) and TopBrain [13]. To extend c… view at source ↗

**Figure 5.** Figure 5: (a) Plugin interface within 3D Slicer. (b) Segmentation and classification results on an MRA scan. 4 Discussion Limitations. First, the challenge provides only center-point aneurysm annotations, so segmentation evaluation remains qualitative ( [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

read the original abstract

Intracranial aneurysms are often asymptomatic until rupture, which carries high mortality. Rupture risk assessment and treatment planning depend on both aneurysm morphology and anatomical location, yet existing automated methods remain limited to binary detection without fine-grained anatomical classification or multi-class segmentation. We present a multi-task framework that simultaneously performs multi-label classification, multi-class aneurysm segmentation, and multi-class vessel segmentation across 13 anatomical locations and four imaging modalities (CTA, MRA, T2, T1-post). Our two-stage approach combines a fast 2D tri-axial Region of Interest (ROI) extraction method with a 3D multi-task nnU-Net backbone. A dual-decoder design mitigates the extreme volume imbalance between aneurysm and vessel classes, while cross-attention pooling and modality-specific auxiliary heads improve feature learning across heterogeneous inputs. Our two-fold ensemble achieved 2nd place in the RSNA 2025 Intracranial Aneurysm Detection challenge. Code, model weights, and a 3D Slicer plugin are publicly available.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

They landed 2nd in the RSNA aneurysm challenge with an ensemble of a modified nnU-Net, but the abstract gives no ablations so the design choices stay uncredited.

read the letter

The paper's clearest result is a second-place finish in the RSNA 2025 Intracranial Aneurysm Detection challenge using a two-fold ensemble. The method adds a tri-axial 2D ROI extractor in front of a 3D nnU-Net, then uses dual decoders to separate aneurysm and vessel segmentation, plus cross-attention pooling and modality-specific heads for CTA, MRA, T2, and T1-post across 13 locations.

The practical parts are the public code, model weights, and 3D Slicer plugin. Those make the work immediately usable for anyone who wants to reproduce or extend the pipeline. The multi-task framing also matches the clinical need for both location classification and morphology segmentation in one pass.

The main gap is the absence of ablation numbers or single-model scores. The abstract states the ranking but does not show whether the tri-axial ROI, dual-decoder, or cross-attention steps actually moved the needle versus a plain nnU-Net plus ensemble. Without those controls it is hard to separate the architectural contribution from challenge-specific tuning.

This is a methods paper aimed at the medical imaging and neuroradiology crowd. Readers working on similar multi-modal aneurysm or vessel tasks will find the released tools and the handling of class imbalance useful. It is not aimed at a broad CV audience.

The work is solid enough on its own terms to go to peer review, provided the full paper adds the missing ablations and dataset details. The competition result plus open release already gives it a concrete anchor.

Referee Report

1 major / 1 minor

Summary. The manuscript presents a multi-task framework for intracranial aneurysm multi-label classification, multi-class aneurysm segmentation, and multi-class vessel segmentation across 13 anatomical locations and four modalities (CTA, MRA, T2, T1-post). The approach uses a two-stage pipeline consisting of fast 2D tri-axial ROI extraction followed by a 3D multi-task nnU-Net backbone with dual decoders to address class imbalance, cross-attention pooling, and modality-specific auxiliary heads. The central empirical claim is that a two-fold ensemble of this model achieved 2nd place in the RSNA 2025 Intracranial Aneurysm Detection challenge.

Significance. If the architectural modifications can be shown to drive the ranking beyond ensembling, the framework would advance automated aneurysm analysis by providing simultaneous classification and fine-grained segmentation that existing binary-detection methods lack. The public release of code, model weights, and a 3D Slicer plugin is a clear strength that supports reproducibility and downstream use.

major comments (1)

[Abstract] Abstract: The 2nd-place ranking is reported for a two-fold ensemble, yet no ablation results, single-model baselines, or direct comparisons against plain nnU-Net on the same data splits are provided to establish that the tri-axial ROI extraction, dual-decoder design, cross-attention pooling, or modality-specific heads are the primary contributors rather than the ensemble procedure or challenge-specific tuning.

minor comments (1)

[Abstract] Abstract: Dataset size, class distribution, exclusion criteria, and any error bars or statistical details around the challenge ranking are omitted, limiting assessment of result robustness.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed review and constructive feedback. We address the major comment below and commit to revisions that strengthen the empirical validation of our contributions.

read point-by-point responses

Referee: [Abstract] Abstract: The 2nd-place ranking is reported for a two-fold ensemble, yet no ablation results, single-model baselines, or direct comparisons against plain nnU-Net on the same data splits are provided to establish that the tri-axial ROI extraction, dual-decoder design, cross-attention pooling, or modality-specific heads are the primary contributors rather than the ensemble procedure or challenge-specific tuning.

Authors: We agree that the current manuscript does not include explicit ablation studies or single-model baselines on the challenge data splits, which limits the ability to isolate the contribution of each proposed component from the ensemble procedure. In the revised manuscript we will add (1) performance metrics for the single-model (non-ensembled) version, (2) a direct comparison against a plain nnU-Net baseline trained on the same splits, and (3) targeted ablations removing the tri-axial ROI extractor, dual-decoder, and modality-specific heads. These additions will be reported on the official validation set to clarify the source of the ranking improvement. revision: yes

Circularity Check

0 steps flagged

Empirical challenge result with no derivation chain or self-referential reductions

full rationale

The paper reports an empirical 2nd-place ranking on the external RSNA 2025 hidden test set using a two-fold ensemble of a multi-task nnU-Net variant. No equations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The central claim is a factual external ranking rather than a derived quantity that reduces to the method's own inputs by construction. The architecture descriptions (tri-axial ROI, dual-decoder, cross-attention) are presented as design choices whose value is assessed by challenge performance, not by internal redefinition or ansatz smuggling.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The work rests on the established effectiveness of nnU-Net for medical segmentation and on standard deep-learning assumptions about multi-task learning and attention mechanisms; no new physical constants, mathematical axioms, or postulated entities are introduced.

free parameters (2)

nnU-Net training hyperparameters
Learning rate, patch size, augmentation parameters, and loss weights are tuned on the challenge training data to achieve the reported ranking.
tri-axial ROI extraction parameters
Thresholds or heuristics used to define the 2D projections and crop bounds are chosen to suit the dataset.

axioms (2)

domain assumption nnU-Net provides a competitive baseline for volumetric medical image segmentation
The paper adopts nnU-Net as backbone without re-deriving or re-validating its performance on this task.
domain assumption Dual-decoder and cross-attention designs mitigate class imbalance and modality heterogeneity
These modifications are presented as solutions without quantitative isolation of their contribution in the abstract.

pith-pipeline@v0.9.1-grok · 5741 in / 1628 out tokens · 41095 ms · 2026-06-26T05:42:02.039113+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

14 extracted references

[1]

Patterns2(2), 100197 (2021) 2 Intracranial Aneurysm Classification and Segmentation 11

Bo, Z.H., Qiao, H., Tian, C., Guo, Y., Li, W., Liang, T., Li, D., Liao, D., Zeng, X., Mei, L., et al.: Toward human intervention-free clinical diagnosis of intracranial aneurysm via deep neural network. Patterns2(2), 100197 (2021) 2 Intracranial Aneurysm Classification and Segmentation 11

2021
[2]

Neu- roinformatics21(1), 21–34 (2023) 2

Di Noto, T., Marie, G., Tourbier, S., Aleman-Gomez, Y., Esteban, O., Saliou, G., Cuadra, M.B., Hagmann, P., Richiardi, J.: Towards automated brain aneurysm detection in TOF-MRA: open data, weak labels, and anatomical knowledge. Neu- roinformatics21(1), 21–34 (2023) 2

2023
[3]

Medical Image Analysis101, 103493 (2025) 1

Hsu, W.C., Meuschke, M., Frangi, A.F., Preim, B., Lawonn, K.: A survey of in- tracranial aneurysm detection and segmentation. Medical Image Analysis101, 103493 (2025) 1

2025
[4]

Nature methods18(2), 203–211 (2021) 2, 3, 5, 8

Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods18(2), 203–211 (2021) 2, 3, 5, 8

2021
[5]

In: International Conference on Medical Image Computing and Computer- Assisted Intervention

Isensee, F., Wald, T., Ulrich, C., Baumgartner, M., Roy, S., Maier-Hein, K., Jaeger, P.F.: nnu-net revisited: A call for rigorous validation in 3d medical image segmen- tation. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 488–498. Springer (2024) 2

2024
[6]

Rudie, J., Calabrese, E., Ball, R., Chang, P., Chen, R., Colak, E., de Verdier, M.C., Prevedello, L., Richards, T., Saluja, R., Zaharchuk, G., Sho, J., Vazirabad,M.:RSNA2025intracranialaneurysmdetection.https://kaggle.com/ competitions/rsna-intracranial-aneurysm-detection(2025), kaggle 2, 8

2025
[7]

Nature communications11(1), 6090 (2020) 2

Shi, Z., Miao, C., Schoepf, U.J., Savage, R.H., Dargis, D.M., Pan, C., Chai, X., Li, X.L., Xia, S., Zhang, X., et al.: A clinically applicable deep-learning model for detecting intracranial aneurysm in computed tomography angiography images. Nature communications11(1), 6090 (2020) 2

2020
[8]

Neuroimage238, 118216 (2021) 1

Timmins, K.M., van der Schaaf, I.C., Bennink, E., Ruigrok, Y.M., An, X., Baum- gartner, M., Bourdon, P., De Feo, R., Di Noto, T., Dubost, F., et al.: Comparing methods of detecting and segmenting unruptured intracranial aneurysms on tof- mras: the adam challenge. Neuroimage238, 118216 (2021) 1

2021
[9]

Radiology290(1), 187–194 (2019) 2

Ueda, D., Yamamoto, A., Nishimori, M., Shimono, T., Doishita, S., Shimazaki, A., Katayama, Y., Fukumoto, S., Choppin, A., Shimahara, Y., et al.: Deep learning for mr angiography: automated detection of cerebral aneurysms. Radiology290(1), 187–194 (2019) 2

2019
[10]

Radiology: Artificial Intelligence 5(5), e230024 (2023) 7, 9

Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., et al.: Totalsegmentator: robust segmen- tation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence 5(5), e230024 (2023) 7, 9

2023
[11]

Radiology312(2), e233197 (2024) 2

Wei, J., Song, X., Wei, X., Yang, Z., Dai, L., Wang, M., Sun, Z., Jin, Y., Ma, C., Hu, C., et al.: Knowledge-augmented deep learning for segmenting and detecting cerebral aneurysms with ct angiography: a multicenter study. Radiology312(2), e233197 (2024) 2

2024
[12]

ArXiv pp

Yang, K., Musio, F., Ma, Y., Juchler, N., Paetzold, J.C., Al-Maskari, R., Höher, L., Li, H.B., Hamamci, I.E., Sekuboyina, A., et al.: Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra. ArXiv pp. arXiv–2312 (2025) 4, 7

2025
[13]

medRxiv pp

Yang, K., Shi, P., Huang, H., Musio, F., Baazaoui, H., Aydin, O.U., Hilbert, A., Hamadache, R.E., Yalcin, C., Zhang, M., et al.: Topbrain segmentation challenge for whole brain vessel anatomy. medRxiv pp. 2026–05 (2026) 4, 7

2026
[14]

IEEE Transactions on Medical Imaging 44(3), 1273–1283 (2024) 2

Yao, L., Chen, D., Zhao, X., Fei, M., Song, Z., Xue, Z., Zhan, Y., Song, B., Shi, F., Wang, Q., et al.: Aaseg: Artery-aware global-to-local framework for aneurysm segmentation in head and neck cta images. IEEE Transactions on Medical Imaging 44(3), 1273–1283 (2024) 2

2024

[1] [1]

Patterns2(2), 100197 (2021) 2 Intracranial Aneurysm Classification and Segmentation 11

Bo, Z.H., Qiao, H., Tian, C., Guo, Y., Li, W., Liang, T., Li, D., Liao, D., Zeng, X., Mei, L., et al.: Toward human intervention-free clinical diagnosis of intracranial aneurysm via deep neural network. Patterns2(2), 100197 (2021) 2 Intracranial Aneurysm Classification and Segmentation 11

2021

[2] [2]

Neu- roinformatics21(1), 21–34 (2023) 2

Di Noto, T., Marie, G., Tourbier, S., Aleman-Gomez, Y., Esteban, O., Saliou, G., Cuadra, M.B., Hagmann, P., Richiardi, J.: Towards automated brain aneurysm detection in TOF-MRA: open data, weak labels, and anatomical knowledge. Neu- roinformatics21(1), 21–34 (2023) 2

2023

[3] [3]

Medical Image Analysis101, 103493 (2025) 1

Hsu, W.C., Meuschke, M., Frangi, A.F., Preim, B., Lawonn, K.: A survey of in- tracranial aneurysm detection and segmentation. Medical Image Analysis101, 103493 (2025) 1

2025

[4] [4]

Nature methods18(2), 203–211 (2021) 2, 3, 5, 8

Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods18(2), 203–211 (2021) 2, 3, 5, 8

2021

[5] [5]

In: International Conference on Medical Image Computing and Computer- Assisted Intervention

Isensee, F., Wald, T., Ulrich, C., Baumgartner, M., Roy, S., Maier-Hein, K., Jaeger, P.F.: nnu-net revisited: A call for rigorous validation in 3d medical image segmen- tation. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 488–498. Springer (2024) 2

2024

[6] [6]

Rudie, J., Calabrese, E., Ball, R., Chang, P., Chen, R., Colak, E., de Verdier, M.C., Prevedello, L., Richards, T., Saluja, R., Zaharchuk, G., Sho, J., Vazirabad,M.:RSNA2025intracranialaneurysmdetection.https://kaggle.com/ competitions/rsna-intracranial-aneurysm-detection(2025), kaggle 2, 8

2025

[7] [7]

Nature communications11(1), 6090 (2020) 2

Shi, Z., Miao, C., Schoepf, U.J., Savage, R.H., Dargis, D.M., Pan, C., Chai, X., Li, X.L., Xia, S., Zhang, X., et al.: A clinically applicable deep-learning model for detecting intracranial aneurysm in computed tomography angiography images. Nature communications11(1), 6090 (2020) 2

2020

[8] [8]

Neuroimage238, 118216 (2021) 1

Timmins, K.M., van der Schaaf, I.C., Bennink, E., Ruigrok, Y.M., An, X., Baum- gartner, M., Bourdon, P., De Feo, R., Di Noto, T., Dubost, F., et al.: Comparing methods of detecting and segmenting unruptured intracranial aneurysms on tof- mras: the adam challenge. Neuroimage238, 118216 (2021) 1

2021

[9] [9]

Radiology290(1), 187–194 (2019) 2

Ueda, D., Yamamoto, A., Nishimori, M., Shimono, T., Doishita, S., Shimazaki, A., Katayama, Y., Fukumoto, S., Choppin, A., Shimahara, Y., et al.: Deep learning for mr angiography: automated detection of cerebral aneurysms. Radiology290(1), 187–194 (2019) 2

2019

[10] [10]

Radiology: Artificial Intelligence 5(5), e230024 (2023) 7, 9

Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., et al.: Totalsegmentator: robust segmen- tation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence 5(5), e230024 (2023) 7, 9

2023

[11] [11]

Radiology312(2), e233197 (2024) 2

Wei, J., Song, X., Wei, X., Yang, Z., Dai, L., Wang, M., Sun, Z., Jin, Y., Ma, C., Hu, C., et al.: Knowledge-augmented deep learning for segmenting and detecting cerebral aneurysms with ct angiography: a multicenter study. Radiology312(2), e233197 (2024) 2

2024

[12] [12]

ArXiv pp

Yang, K., Musio, F., Ma, Y., Juchler, N., Paetzold, J.C., Al-Maskari, R., Höher, L., Li, H.B., Hamamci, I.E., Sekuboyina, A., et al.: Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra. ArXiv pp. arXiv–2312 (2025) 4, 7

2025

[13] [13]

medRxiv pp

Yang, K., Shi, P., Huang, H., Musio, F., Baazaoui, H., Aydin, O.U., Hilbert, A., Hamadache, R.E., Yalcin, C., Zhang, M., et al.: Topbrain segmentation challenge for whole brain vessel anatomy. medRxiv pp. 2026–05 (2026) 4, 7

2026

[14] [14]

IEEE Transactions on Medical Imaging 44(3), 1273–1283 (2024) 2

Yao, L., Chen, D., Zhao, X., Fei, M., Song, Z., Xue, Z., Zhan, Y., Song, B., Shi, F., Wang, Q., et al.: Aaseg: Artery-aware global-to-local framework for aneurysm segmentation in head and neck cta images. IEEE Transactions on Medical Imaging 44(3), 1273–1283 (2024) 2

2024