Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

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

TL;DR

This work tackles pediatric brain tumor segmentation in mpMRI within BraTS 2025 Task-6 (PED) by extending the nnU-Net framework with targeted, task-specific enhancements. Key innovations include a widened residual encoder with SE attention, 3D depthwise separable convolutions, a specificity-driven regularization term, and small-scale Gaussian initialization, complemented by two postprocessing steps. The proposed model achieves state-of-the-art lesion-wise Dice on the Task-6 validation, with CC $0.759$, ED $0.967$, ET $0.826$, NET $0.910$, TC $0.928$, and WT $0.928$, demonstrating robust performance across pediatric tumor subregions. The authors note remaining challenges in ET and CC precision and suggest future exploration of attention mechanisms and Transformer-based approaches to further improve pediatric tumor segmentation.

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).

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

TL;DR

This work tackles pediatric brain tumor segmentation in mpMRI within BraTS 2025 Task-6 (PED) by extending the nnU-Net framework with targeted, task-specific enhancements. Key innovations include a widened residual encoder with SE attention, 3D depthwise separable convolutions, a specificity-driven regularization term, and small-scale Gaussian initialization, complemented by two postprocessing steps. The proposed model achieves state-of-the-art lesion-wise Dice on the Task-6 validation, with CC , ED , ET , NET , TC , and WT , demonstrating robust performance across pediatric tumor subregions. The authors note remaining challenges in ET and CC precision and suggest future exploration of attention mechanisms and Transformer-based approaches to further improve pediatric tumor segmentation.

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).

Paper Structure

This paper contains 16 sections, 4 equations, 8 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Model Architecture
  3. Baseline: standard nnU-Net(v2)
  4. Widened residual encoder with SE Attention
  5. Depthwise Separable Convolutions
  6. Standard Convolution
  7. Depthwise Convolution
  8. Pointwise Convolution
  9. Depthwise Separable Convolution
  10. Regularization
  11. Initialization
  12. Postprocessing
  13. Training
  14. Results
  15. Discussion
  16. ...and 1 more sections

Figures (8)

  • Figure 1: Graphical representation of data processing and annotations in pediatric brain tumors. Top panel presents the processing pipeline, and the bottom panel illustrates the annotated tumor subregions along with mpMRI structural scans (T1, T1CE, T2, and T2-FLAIR). Tumor subregions include the enhancing tumor (ET - red), non-enhancing tumor (NET - green), cystic component (CC - yellow), and edema (ED - teal) regions.
  • Figure 2: Overview of our enhanced nnU-Net. The left branch is the encoder (downsampling), the right branch is the decoder (upsampling), and dashed arrows denote skip connections.
  • Figure 3: Left: Standard convolution with norm and LeakyReLU. Right: Depthwise Separable convolutions with norm and LeakyReLU.
  • Figure 4: Standard Convolution
  • Figure 5: Depthwise and Pointwise Convolution
  • ...and 3 more figures