Overview

Schema Version

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

Name

Longitudinal assessment of posttreatment diffuse glioma with 3D nnU-Net (segmentation and longitudinal change networks)

Link

https://pubs.rsna.org/doi/10.1148/ryai.210243

Indexing

Keywords: nnU-Net, U-Net, posttreatment glioma, longitudinal assessment, segmentation, enhancing tumor, edema, necrotic core, RANO

Content: MR, NR, OI

RadLex: RID35976, RID4027, RID10795, RID35806, RID49779, RID7106, RID49531, RID10794

SNOMED: 1157043006, 443936004, 1163375002, 1156974002

Author(s)

Jeffrey D. Rudie

Evan Calabrese

Rachit Saluja

David Weiss

John B. Colby

Soonmee Cha

Christopher P. Hess

Andreas M. Rauschecker

Leo P. Sugrue

Javier E. Villanueva-Meyer

Organization(s)

University of California, San Francisco, Department of Radiology and Biomedical Imaging

Hospital of the University of Pennsylvania, Department of Radiology

Version

1.0

License

URL: https://pubs.rsna.org/doi/10.1148/ryai.210243

Contact

Jeffrey D. Rudie, MD, PhD, email: Jeff.Rudie@gmail.com

Funding

Supported in part by an American Society of Neuroradiology Foundation Grant in Artificial Intelligence.

Ethical review

Institutional review board approved, HIPAA-compliant; waiver of written consent.

Date

Updated: 2022-09-01

Published: 2022-08-03

Created: 2021-09-14

References

[1] Rudie JD; Calabrese E; Saluja R; Weiss D; Colby JB; Cha S; Hess CP; Rauschecker AM; Sugrue LP; Villanueva-Meyer JE. "Longitudinal Assessment of Posttreatment Diffuse Glioma Tissue Volumes with Three-dimensional Convolutional Neural Networks". Radiology: Artificial Intelligence. 2022;4(5):e210243. 2022-08-03. doi:10.1148/ryai.210243. PMID: 36204543. PMCID: PMC9530762.

Model

Architecture

Three-dimensional nnU-Net convolutional neural networks (self-configuring U-Net variants) with multichannel inputs; separate networks for segmentation and for longitudinal change (ED and AT).

Clinical benefit

Automated quantification and localization of posttreatment diffuse glioma tissue subregions and their longitudinal changes to support routine follow-up assessment and clinical trials.

Clinical workflow phase

Clinical decision support systems; workflow optimization for longitudinal tumor burden assessment.

Decision threshold

Tumor segmentation-based classification thresholds: ED increased/decreased at +15%/−15%; AT increased/decreased at +15%/−20% with minimum AT change ≥0.5 cm3. Longitudinal change networks: ED net +0.2 cm3 / −0.5 cm3; AT net +0.1 cm3 / −0.25 cm3 for increased/decreased.

Degree of automation

Fully automated preprocessing, segmentation, and longitudinal change localization/classification using nnU-Net defaults.

Indications for use

Posttreatment patients with diffuse glioma undergoing routine follow-up brain MRI; intended to segment ED, AT, and NCR and to classify increased/decreased/unchanged changes in ED and AT between two consecutive time points.

Input

3D brain MRI sequences: T1-weighted precontrast, T1-weighted postcontrast, T2-weighted, and FLAIR. Longitudinal change networks use registered time 1 and time 2 images and subtraction images (FLAIR for ED; postcontrast–precontrast T1 subtraction images for AT) as four-channel inputs.

Instructions

Automated pipeline including intermodality registration, resampling to 1×1×1 mm, skull stripping, and bias correction; for longitudinal models, register across time points and form subtraction images prior to inference.

Limitations

Single-site, single-vendor 3.0-T scanners; voxelwise annotations by a single expert; trained and tested on a single-institution cohort which may limit generalizability. Designed to quantify volume changes in ED and AT, not to determine true progression vs treatment-related effects (e.g., pseudoprogression).

Output

CDEs: RDE2038, RDE1959, RDE1955, RDE1956, RDE1958, RDE1957, RDE1960

Description: Voxelwise segmentation masks for ED, AT, and NCR at individual time points; localized maps of increasing/decreasing ED and AT across time points; categorical classification of longitudinal change (increased, decreased, unchanged) for ED and AT.

Recommendation

Demonstrated performance comparable to neuroradiologists for longitudinal change classification; further multisite validation recommended before clinical deployment.

Reproducibility

Implemented with default nnU-Net self-configuration for preprocessing, architecture, and hyperparameters; details provided in supplemental appendices (E3–E5).

Use

Intended: Image segmentation, Detection and diagnosis

Out-of-scope: Exam protocol selection, Detection and diagnosis

Excluded: Detection and diagnosis

User

Intended: Radiologist, Subspecialist diagnostic radiologist, Researcher

Out-of-scope: Patient, Layperson

Excluded: Non-physician provider