Towards Reliable Pediatric Brain Tumor Segmentation: Task-Specific nnU-Net Enhancements

Xiaolong Li, Zhi‐Qin John Xu, Yan Ren, Tianming Qiu, Xiaowen Wang

Abstract

Accurate segmentation of pediatric brain tumors in multi-parametric magnetic resonance imaging (mpMRI) is critical for diagnosis, treatment planning, and monitoring, yet faces unique challenges due to limited data, high anatomical variability, and heterogeneous imaging across institutions. In this work, we present an advanced nnU-Net framework tailored for BraTS 2025 Task-6 (PED), the largest public dataset of pre-treatment pediatric high-grade gliomas. Our contributions include: (1) a widened residual encoder with squeeze-and-excitation (SE) attention; (2) 3D depthwise separable convolutions; (3) a specificity-driven regularization term; and (4) small-scale Gaussian weight initialization. We further refine predictions with two postprocessing steps. Our models achieved first place on the Task-6 validation leaderboard, attaining lesion-wise Dice scores of 0.759 (CC), 0.967 (ED), 0.826 (ET), 0.910 (NET), 0.928 (TC) and 0.928 (WT).

Introduction

Automated segmentation of pediatric brain tumors is a high-stakes challenge in neuro-oncology. While tools like the standard nnU-Net have revolutionized medical imaging, pediatric cases present unique hurdles. These include significant data scarcity compared to adult cohorts, high anatomical variability across different developmental stages, and the frequent absence of specific tumor sub-components (e.g., the "enhancing tumor" or ET). Standard neural networks often struggle to predict a "perfect zero" for absent classes, leading to false positives that significantly degrade lesion-wise Dice scores.

Overview of the data processing and tumor segmentation pipeline, including data sources and annotation classes.

This study addresses these challenges within the BraTS 2025 Task-6 (PED) framework, utilizing the largest public collection of mpMRI scans for pediatric high-grade gliomas, which include T1, T1-contrast enhanced (T1CE), T2, and T2-FLAIR modalities.

Methodology

The proposed "Advanced nnU-Net" framework builds upon the established nnU-Net (v2) baseline with several task-specific architectural and algorithmic enhancements:

Widened Residual Encoder

Unlike the symmetric design of standard U-Nets, the authors implemented an asymmetric architecture by doubling the number of feature channels in the encoder relative to the decoder. This allows the network to capture a richer set of spatial and textural features before spatial downsampling.

The proposed widened residual encoder architecture featuring depthwise separable convolutions, squeeze-and-excitation blocks, and skip connections.

Efficiency and Attention

To maintain a manageable parameter count despite the widened encoder, the framework utilizes 3D Depthwise Separable Convolutions. Furthermore, Squeeze-and-Excitation (SE) blocks are integrated to provide channel-wise attention, enabling the model to prioritize the most relevant feature channels for segmentation.

Regularization and Post-processing

To address the "false positive" problem for absent tumor classes, a novel specificity-driven regularization term was added to the loss function. This term explicitly penalizes the model for predicting a class when it should not exist. Additionally, the team leveraged biological priors by using the T1CE/T1 intensity ratio in post-processing to refine final voxel labels, effectively differentiating enhancing from non-enhancing regions.

Results and Discussion

The enhanced framework secured first place on the BraTS-PED validation leaderboard. Performance was evaluated using lesion-wise Dice scores across several critical subregions:

  • Edema (ED): 0.967
  • Whole Tumor (WT): 0.928
  • Tumor Core (TC): 0.928
  • Non-Enhancing Tumor (NET): 0.910
  • Enhancing Tumor (ET): 0.826
  • Cystic Component (CC): 0.759

A key insight from the ablation studies was that doubling the encoder width led to a consistent 2–4% improvement in Dice scores, suggesting that feature extraction is the primary bottleneck in complex pediatric cases. While the model demonstrated high precision for the overall tumor volume, the Cystic Component (CC) remains the most difficult to segment accurately due to its high variability.

Conclusion

By tailoring the nnU-Net architecture with a widened encoder, attention mechanisms, and specificity-driven loss, this work establishes a new benchmark for reliable pediatric brain tumor segmentation. Future work will focus on more sophisticated attention mechanisms and finding ways to better exploit limited pediatric datasets where data-hungry architectures like Transformers currently struggle.

Summary updated

Abstract

Accurate segmentation of pediatric brain tumors in multi-parametric magnetic resonance imaging (mpMRI) is critical for diagnosis, treatment planning, and monitoring, yet faces unique challenges due to limited data, high anatomical variability, and heterogeneous imaging across institutions. In this work, we present an advanced nnU-Net framework tailored for BraTS 2025 Task-6 (PED), the largest public dataset of pre-treatment pediatric high-grade gliomas. Our contributions include: (1) a widened residual encoder with squeeze-and-excitation (SE) attention; (2) 3D depthwise separable convolutions; (3) a specificity-driven regularization term; and (4) small-scale Gaussian weight initialization. We further refine predictions with two postprocessing steps. Our models achieved first place on the Task-6 validation leaderboard, attaining lesion-wise Dice scores of 0.759 (CC), 0.967 (ED), 0.826 (ET), 0.910 (NET), 0.928 (TC) and 0.928 (WT).
BibTeX
@misc{li2025towards,
  author = {Xiaolong Li and Zhi‐Qin John Xu and Yan Ren and Tianming Qiu and Xiaowen Wang},
  title = {{Towards Reliable Pediatric Brain Tumor Segmentation: Task-Specific nnU-Net Enhancements}},
  year = {2025},
  eprint = {2511.00449},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.00449},
}
Figure