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

nnU-Net–based segmentation of pediatric medulloblastoma tumor subcompartments (nnU-NetTL and nnU-NetDL)

https://doi.org/10.1148/ryai.230115

1.0

Corresponding author: Pallavi Tiwari (email: ude.csiw@9irawitp)

Supported by Musella Foundation for Brain Tumor Research and Information; R&D Pilot Award, Departments of Radiology and Medical Physics, University of Wisconsin–Madison; National Cancer Institute (1U01CA248226, 1R01CA264017, 3U01CA248226-S1, 1R01CA277728, P30CA014520); Department of Defense PRCRP (W81XWH-18-1-0404); Department of Veterans Affairs (1 I01 BX005842); Dana Foundation David Mahoney Neuroimaging Program; V Foundation Translational Research Award; Johnson & Johnson WiSTEM2D Award; NCATS/UL1TR002373; Wisconsin Alumni Research Foundation WARF Accelerator Oncology Diagnostics Grant (MSN281757).

Retrospective, HIPAA-compliant; IRB approval at each site (ID 2022–1683); informed consent waived.

nnU-Net (self-configuring U-Net) with automated configuration selection among 2D U-Net, 3D U-Net, and 3D U-Net cascade; transfer learning variant fine-tuned using Models Genesis from a source model pretrained on adult glioma (BraTS).

MRI data from hospitals A and B under institutional agreements; hospital C data via CBTN. Segmentations from CBTN studies will be released into the CBTN network. No public code repository listed.

Automated delineation of medulloblastoma tumor habitat and subcompartments to support accurate volumetrics and potentially improve radiation therapy planning and response assessment.

Clinical decision support systems; treatment planning support; research workflow.

Fully automated segmentation pipeline after required preprocessing (registration to age-specific atlas, skull stripping, bias correction, intensity matching).

Research-use automated segmentation of pediatric medulloblastoma on multiparametric brain MRI (axial Gd-T1w, T2w, and FLAIR) in patients aged 2–18 years in hospital/clinical imaging environments.

Multiparametric brain MRI: gadolinium-enhanced T1-weighted, T2-weighted, and FLAIR sequences; preprocessed with age-specific atlas registration, skull stripping, bias correction, and intensity matching.

Use axial Gd-T1w, T2w, and FLAIR MRI. Preprocess by registering to an age-appropriate atlas (0–2, >2–5, >5–10, >10–18 years), skull stripping (FSL BET), N4 bias correction, and intensity matching. Train with fivefold cross-validation; nnU-Net selects best configuration/ensemble. For transfer learning, pretrain on adult glioma (BraTS), then fine-tune all layers using Models Genesis on pediatric data.

Small multi-institutional cohort (n=78); challenges in precise reference labels due to pediatric imaging constraints and tumor heterogeneity; difficulty distinguishing ET vs NET when intensities are similar; ED subcompartment often absent, ED labels augmented around ventricles; CC was scarce and combined with NET; poor performance for very small lesion regions; occasional over/under-segmentation; potential skull stripping errors; generalization beyond studied sites and imaging protocols not established.

Intended for research use; prospective validation and regulatory clearance required before clinical deployment.

Fivefold cross-validation; robustness assessed via three train/test site splits (train on two sites, test on third) with consistent Dice scores across sites; preprocessing and model configurations described to enable replication.