Fast 2D tri-axial ROI extraction + 3D Multi-Task Segmentation and Classification for Intracranial Aneurysms

2026-08-0694596442

Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

Fast 2D tri-axial ROI extraction + 3D Multi-Task Segmentation and Classification for Intracranial Aneurysms

Link

https://github.com/PengchengShi1220/RSNA2025_Intracranial-Aneurysm-Detection

Indexing

Keywords: intracranial aneurysm, segmentation, classification, deep learning, nnU-Net, ROI extraction, multi-task learning, vascular segmentation

Content: CT, HN, MR, NR, OI

RadLex: RID28600, RID4710

SNOMED: 128609009

Author(s)

Pengcheng Shi

Yan Lu

Jiawei Chen

Kaiyuan Yang

Houjing Huang

Organization(s)

Medical Image Insights, Shanghai

University of Zurich (UZH)

Version

1.0

License

Text: Apache 2.0

URL: https://www.apache.org/licenses/LICENSE-2.0

Funding

Bjoern Menze and the Helmut Horten Foundation provided funding support.

Date

Published: 2025-10-15

Model

Architecture

Multi-stage deep learning approach based on nnU-Net framework including a fast 2D tri-axial ROI extraction model and a 3D multi-task nnU-Net model with cross-attention pooling and modality classification heads.

Availability

Code and models are available on GitHub at https://github.com/PengchengShi1220/RSNA2025_Intracranial-Aneurysm-Detection and on Kaggle model repository.

Clinical benefit

Automated detection and classification of intracranial aneurysms assisting radiological diagnosis from CT and MRI scans.

Clinical workflow phase

Clinical decision support system, assisting in interpretation and classification during image reading phase.

Decision threshold

Not explicitly specified; classification outputs are probabilistic with model ensembles and TTA used for robust predictions.

Degree of automation

Supports decision-making by automatically segmenting vessels and aneurysms and providing classification outputs; does not fully automate diagnosis.

Indications for use

For detection and classification of intracranial aneurysms in patients undergoing vascular imaging with CT or MRI.

Input

Preprocessed CT and MRI volumes resized to 224x224x224 including vascular regions extracted via 2D tri-axial segmentation.

Instructions

Use two-stage approach: first perform 2D tri-axial ROI extraction to locate vascular regions, then apply 3D multi-task nnU-Net model for vessel and aneurysm segmentation and classification with heavy test time augmentation and ensembling for best performance.

Limitations

Model trained on data with certain manual corrections; possible sensitivity to imaging variations; limited by Kaggle platform constraints; some data class imbalances addressed by oversampling and weighting.

Output

CDEs: RDE1289, RDE1292, RDE1293, RDE1287, RDE1290.5, RDE1290.1, RDE1290

Description: Outputs include multi-class segmentation masks for vessels and aneurysms, and multi-label classification results indicating the presence of aneurysms across 13 anatomical locations.

Recommendation

Use ensemble predictions with test time augmentation for increased robustness; manual correction of annotations improves segmentation training; model code and checkpoints are publicly available for reproducibility.

Reproducibility

Model training and inference codes are publicly available enabling reproducibility; trained on external and challenge datasets with manual annotation corrections.

Use

Intended: Intracranial aneurysm detection and classification

User

Intended: Radiologists, Clinical researchers, Medical imaging specialists