Overview

Schema Version

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

Name

Fully Automated 3D Vestibular Schwannoma Segmentation (nnU-Net) on Contrast-enhanced T1- and T2-weighted MRI

Link

https://dx.doi.org/10.1148/ryai.210300

Indexing

Keywords: vestibular schwannoma, acoustic neuroma, MRI, T1-weighted contrast-enhanced, T2-weighted, segmentation, nnU-Net, U-Net, deep learning, multicenter, multivendor

Content: MR, NR, HN

RadLex: RID12775, RID9392, RID4473, RID20447, RID49531, RID10796

Author(s)

Olaf M. Neve

Yunjie Chen

Qian Tao

Stephan R. Romeijn

Nick P. de Boer

Willem Grootjans

Mark C. Kruit

Boudewijn P. F. Lelieveldt

Jeroen C. Jansen

Erik F. Hensen

Berit M. Verbist

Marius Staring

Organization(s)

Leiden University Medical Center (Departments of Otorhinolaryngology and Head & Neck Surgery; Division of Image Processing, Department of Radiology; Department of Radiology)

Delft University of Technology, Knowledge Driven AI Lab

Version

1.0

License

Funding

Supported by a strategic fund of the Leiden University Medical Center. Y.C. supported by the China Scholarship Council (grant 202008130140).

Ethical review

Institutional review board approved (Leiden University Medical Center IRB G19.115); informed consent waived.

Date

Published: 2022-06-22

Created: 2021-12-06

References

[1] Neve OM, Chen Y, Tao Q, Romeijn SR, de Boer NP, Grootjans W, Kruit MC, Lelieveldt BPF, Jansen JC, Hensen EF, Verbist BM, Staring M.. "Fully Automated 3D Vestibular Schwannoma Segmentation with and without Gadolinium-based Contrast Material: A Multicenter, Multivendor Study". Radiology: Artificial Intelligence. 2022 Jul;4(4):e210300.. 2022-07-01. doi:10.1148/ryai.210300. PMID: 35923375. PMCID: PMC9344213.

Model

Architecture

3D nnU-Net (3D U-Net with five encoder/decoder layers); multiclass segmentation of intra- and extrameatal components; PyTorch 1.7.1; trained with fivefold cross-validation and ensemble averaging of softmax scores.

Availability

Data available from the corresponding author upon reasonable request (per article).

Clinical benefit

Automated, accurate volumetric measurement and delineation of vestibular schwannomas (whole, intra- and extrameatal parts) on MRI; supports surveillance and treatment decision-making while reducing manual workload.

Clinical workflow phase

Clinical decision support systems; workflow optimization for tumor surveillance and follow-up measurements.

Decision threshold

Tumor detection defined as at least one voxel detected; connected-component postprocessing retains largest component.

Degree of automation

Fully automated 3D segmentation pipeline with automated preprocessing, inference, and postprocessing; supports clinician review.

Indications for use

Automated detection and segmentation of unilateral vestibular schwannoma in adults on contrast-enhanced T1-weighted and/or high-resolution T2-weighted MRI in pre-treatment settings within radiology workflows at tertiary and general hospitals.

Input

MRI volumes: gadolinium-enhanced T1-weighted scans and/or high-resolution T2-weighted scans of the cerebellopontine angle/internal auditory canal region.

Instructions

Provide CE T1-weighted and/or high-resolution T2-weighted MRI. Images are z-score normalized and resampled to median spacing. Inference uses a sliding window (320×320×10 voxels) with Gaussian-weighted aggregation; largest connected component is kept. Outputs include whole tumor and intra-/extrameatal masks and volumes. Clinician review recommended, especially for cystic tumors and atypical cases.

Limitations

Retrospective, multicenter training; T2 ground truth derived from registered T1 delineations; slightly lower performance on T2-weighted images and in polycystic tumors; occasional false positives outside region of interest; trained only on pre-treatment scans—does not generalize to post-surgery or post-irradiation without retraining.

Output

CDEs: RDE1955, RDE1278, RDE2042

Description: 3D segmentation masks for vestibular schwannoma, partitioned into intra- and extrameatal components, with derived volume measurements and detection status.

Recommendation

Use primarily on pre-treatment CE T1 and/or high-resolution T2 MRI for automated vestibular schwannoma volumetry; verify results in cases with large peripheral cysts or atypical appearance; not intended for post-treatment follow-up without retraining.

Regulatory information

Authorization status: Research-use only; not FDA- or CE-cleared as a medical device per article.

Reproducibility

Fivefold cross-validation with ensemble of five models; robust performance on an independent multicenter test set and an external public dataset; median runtime ~78 seconds per patient.

Sustainability

Training on NVIDIA Tesla V100 16 GB; median inference time ~78 s per patient; computational footprint typical for 3D nnU-Net.

Use

Intended: Image segmentation, Detection and diagnosis

Out-of-scope: Detection and diagnosis

Excluded: Treatment

User

Intended: Radiology technologist, Referring physician, Subspecialist diagnostic radiologist