Abstract
Accurate segmentation of pediatric brain tumors in multiparametric magnetic resonance imaging (mpMRI) is essential for treatment planning and response assessment, yet faces distinctive challenges compared to adult glioma segmentation due to heterogeneous imaging phenotypes and limited pediatric datasets. We present our submission to BraTS-GoAT 2026 Task 3 (Generalizability Across Tumors), targeting, as a single-population baseline within this multi-population challenge, the segmentation of pediatric high-grade gliomas into necrotic core (NCR), peritumoral edema (ED), and enhancing tumor (ET) on the BraTS Pediatric Tumors dataset (1,351 cases; 894 training, 451 held-out test). Our approach leverages nnU-Net v2 in its 3D full-resolution configuration with targeted engineering optimizations: (1) reduction of parallel data-augmentation workers to mitigate memory constraints on resource-limited clusters; (2) disabling PyTorch dynamic compilation to ensure training stability over extended epochs; and (3) ensemble aggregation with test-time augmentation to maximize generalization. Training on four MRI sequences (T1n, T1c, T2w, T2f) on NVIDIA A10G GPUs within strict computational constraints resulted in 5-fold cross-validation Dice scores of 0.824–0.845 (mean 0.831), with strong performance on enhancing tumor (0.861–0.878) and edema (0.849–0.860), while necrotic core remained more variable (0.758–0.795), reflecting the pediatric-specific imaging heterogeneity. The complete deployment pipeline was rigorously validated end-to-end, including containerized inference verification on held-out test cases, ensuring readiness for the challenge evaluation. Held-out test set performance and cross-population generalizability within Task 3 remain to be established through the challenge evaluation and future extension of this work, respectively.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_GoAT_004.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/BraTS_GoAT_004_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=ZaYIEJEkrq
BibTex
@InProceedings{NazAha_An_MICCAISAT2026,
author = { Nazir, Ahad Ali AND Ul Jabbar, Absaar},
title = { { An Engineering-optimized, Resource-Constrained nnU-Net v2 Pediatric Baseline for the BraTS-GoAT 2026 Generalizability-Across-Tumors Challenge } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17253},
month = {pending},
page = {pending}
}