Overview

Schema Version

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

Name

Deep Learning Segmentation of Infiltrative and Enhancing Cellular Tumor at Pre- and Posttreatment MRI of Glioblastoma

Link

https://dx.doi.org/10.1148/ryai.230489

Indexing

Keywords: Segmentation, Glioblastoma, Multishell diffusion MRI, Restriction spectrum imaging, nnU-Net, Perfusion MRI, Survival prediction, Posttreatment changes, Recurrent tumor

Content: NR, MR, OI

RadLex: RID4044, RID49531, RID38778

SNOMED: 1163375002

Author(s)

Louis Gagnon

Diviya Gupta

George Mastorakos

Nathan White

Vanessa Goodwill

Carrie R. McDonald

Thomas Beaumont

Christopher Conlin

Tyler M. Seibert

Uyen Nguyen

Jona Hattangadi-Gluth

Santosh Kesari

Jessica D. Schulte

David Piccioni

Kathleen M. Schmainda

Nikdokht Farid

Anders M. Dale

Jeffrey D. Rudie

Organization(s)

University of California San Diego

Cortechs.ai

Pacific Neuroscience Institute and Saint John’s Cancer Institute at Providence Saint John’s Health Center

Medical College of Wisconsin

Version

1.0

License

URL: https://pubs.rsna.org/journal/ryai

Contact

ac.lavalu@9.nongag.siuol

Funding

Salaries for two authors (G.M. and J.D.R.) were partly funded by NIH Small Business Innovation Research grant R44NS120796 awarded to Cortechs.ai. Authors declared no funding for this work otherwise.

Ethical review

Institutional Review Board approval obtained at the University of California San Diego with waiver of informed consent; data were de-identified and collected according to HIPAA.

Date

Published: 2024-08-21

Created: 2023-10-31

References

[1] Gagnon L, Gupta D, Mastorakos G, et al.. "Deep Learning Segmentation of Infiltrative and Enhancing Cellular Tumor at Pre- and Posttreatment Multishell Diffusion MRI of Glioblastoma". Radiology: Artificial Intelligence. 2024;6(5):e230489.. 2024-08-21. doi:10.1148/ryai.230489. PMID: 39166970. PMCID: PMC11427928.

Model

Architecture

nnU-Net (standard 3dfullres configuration), convolutional neural network for 3D biomedical image segmentation.

Availability

Trained on UCSD postoperative data; external testing performed on MCW, LUMIERE, UPenn-GBM, and UCSF-PDGM datasets. Data available from corresponding author by request.

Clinical benefit

Automated segmentation of enhancing and nonenhancing cellular tumor to assist detection of infiltrative tumor, distinction of recurrent/residual tumor from posttreatment changes, and prediction of overall and progression-free survival.

Clinical workflow phase

Clinical decision support systems; workflow optimization for tumor assessment and treatment planning.

Decision threshold

For ROC-based classification, varying thresholds of segmented cellular tumor volume were used; example specificity values reported at 1 mL threshold.

Degree of automation

Fully automated image segmentation with outputs intended to support, not replace, clinician decision making.

Indications for use

Patients with glioblastoma undergoing pre- or postoperative brain MRI; intended for use on multimodal MRI including diffusion (multishell/RSI or ADC) and perfusion (DSC) when available in radiology practice or research settings.

Input

Multimodal MRI: T1-weighted, T1-weighted contrast-enhanced, T2-weighted, FLAIR, RSI cellularity map (multishell diffusion), and DSC-derived cerebral blood volume; alternative inputs evaluated with ADC or without diffusion.

Limitations

Does not incorporate clinical variables into the model; limited size of external postoperative test dataset; specific RSI sequence not widely available; lack of histologic validation of cellular tumor maps at this stage; lower Dice scores for NECT, with under-segmentation observed in external testing.

Output

CDEs: RDE2038, RDE1955, RDE1956

Description: Voxelwise segmentation maps and corresponding volumes of total cellular tumor (TCT), enhancing cellular tumor (ECT), and nonenhancing cellular tumor (NECT).

Use

Intended: Risk assessment, Image segmentation, Detection and diagnosis

User

Intended: Physician, Subspecialist diagnostic radiologist, Researcher