Overview

Schema Version

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

Name

Stepwise Transfer Learning nnU-Net for Pediatric Low-Grade Glioma (pLGG) T2-weighted MRI Segmentation (transfer-encoder model)

Link

https://pubs.rsna.org/doi/10.1148/ryai.230254

Indexing

Keywords: Stepwise Transfer Learning, Pediatric Brain Tumors, MRI Segmentation, Deep Learning, pLGG, nnU-Net

Content: MR, NR, PD, OI

RadLex: RID4071, RID12775, RID10312, RID4412, RID10796

SNOMED: 428954009

Author(s)

Aidan Boyd

Zezhong Ye

Sanjay P. Prabhu

Michael C. Tjong

Yining Zha

Anna Zapaishchykova

Sridhar Vajapeyam

Paul J. Catalano

Hasaan Hayat

Rishi Chopra

Kevin X. Liu

Ali Nabavizadeh

Adam C. Resnick

Sabine Mueller

Daphne A. Haas-Kogan

Hugo J. W. L. Aerts

Tina Y. Poussaint

Benjamin H. Kann

Organization(s)

Artificial Intelligence in Medicine (AIM) Program, Mass General Brigham, Harvard Medical School

Department of Radiation Oncology, Brigham and Women’s Hospital, Dana-Farber Cancer Institute, Harvard Medical School

Department of Radiology, Brigham and Women’s Hospital, Harvard Medical School

Department of Radiology, Boston Children’s Hospital, Harvard Medical School

Department of Biostatistics and Computational Biology, Harvard T.H. Chan School of Public Health

Center for Data-Driven Discovery in Biomedicine (D3b), Children’s Hospital of Philadelphia

Department of Neurosurgery, Children’s Hospital of Philadelphia

Department of Radiology, Perelman School of Medicine, University of Pennsylvania

University of California, San Francisco (Departments of Neurology, Pediatrics, and Neurologic Surgery)

Department of Radiology and Nuclear Medicine, CARIM & GROW, Maastricht University

Version

1.0

License

Contact

B.H.K. (email: ude.dravrah.icfd@nnaK_nimajneB)

Funding

Supported by NIH (U54CA274516, U24CA194354, U01CA190234, U01CA209414, R35CA22052, K08DE030216), NCI SPORE (2P50CA165962), European Research Council (866504), RSNA (RSCH2017), Pediatric Low-Grade Astrocytoma Program at the Pediatric Brain Tumor Foundation, Botha-Chan Low Grade Glioma Consortium, William M. Wood Foundation; additional institutional support noted for individual authors.

Ethical review

Study conducted with local IRB approval and waiver of consent due to use of public datasets and retrospective nature, in accordance with the Declaration of Helsinki.

Date

Updated: 2024-07-10

Published: 2024-07-10

Created: 2023-07-11

References

[1] Boyd A, Ye Z, Prabhu SP, et al.. "Stepwise Transfer Learning for Expert-level Pediatric Brain Tumor MRI Segmentation in a Limited Data Scenario". Radiology: Artificial Intelligence. 2024;6(4):e230254.. 2024-07-10. doi:10.1148/ryai.230254. PMID: 38984985. PMCID: PMC11294948.

Model

Architecture

nnU-Net (U-Net-based convolutional neural network) with fivefold cross-validation and ensemble of folds; stepwise in-domain transfer learning with encoder frozen during final fine-tuning (transfer-encoder).

Availability

Code, trained model, and statistical analysis are publicly available: https://github.com/AIM-KannLab/pLGG_Segmentation

Clinical benefit

Automated whole-tumor segmentation and volumetric measurement for pediatric low-grade glioma on T2-weighted MRI to aid risk stratification, monitoring tumor progression, treatment response assessment, and surgical/radiation planning.

Clinical workflow phase

Clinical decision support systems; workflow optimization via autosegmentation for volumetric assessment; outputs intended for clinician review.

Degree of automation

Fully automated segmentation after preprocessing; clinical review of outputs recommended before decision-making.

Indications for use

Segmentation of whole tumor in pediatric low-grade glioma (grade I–II) on preoperative/pre-treatment brain T2-weighted MRI in pediatric/adolescent patients (approximately 0–25 years) in radiology/oncology care settings.

Input

Preoperative/pre-treatment T2-weighted brain MRI (DICOM converted to NIfTI), preprocessed with bias correction, resampling, registration to pediatric template, and brain extraction.

Instructions

Follow provided preprocessing pipeline (DICOM→NIfTI via dcm2nii; N4 bias correction; resample to 1 mm isotropic; rigid registration to NIHPD pediatric template; HD-BET brain extraction). Apply the trained transfer-encoder nnU-Net ensemble for inference. Review AI-generated masks for clinical acceptability before use.

Limitations

Retrospective study; T2-weighted images only; no subregion segmentation; failures observed in cases with ventricular location, large cystic components, poor image quality/large section thickness leading to empty segmentations, and large heterogeneous lesions causing under-segmentation; model trained and validated on specific cohorts (CBTN, DFCI/BCH); outputs require clinician review.

Output

CDEs: RDE1278, RDE2042

Description: Whole-tumor binary segmentation mask for pLGG on T2-weighted MRI with derived tumor volume; DSC and relative volume difference evaluated against expert references.

Recommendation

Use as an assistive tool for pLGG volumetric assessment; ensure clinician review of segmentations prior to clinical decision-making.

Regulatory information

Authorization status: Not a regulated/cleared medical device; research-use model reported in peer-reviewed study.

Reproducibility

Fivefold cross-validation with ensemble of folds; early stopping (no improvement for 50 epochs; max 1000 epochs); default nnU-Net training settings used; external testing performed; code and trained weights publicly available for replication.

Use

Intended: Image segmentation

Out-of-scope: Image segmentation

Excluded: Decision support

User

Intended: Physician, Radiologist, Subspecialist diagnostic radiologist, Researcher