Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Frequency-Aware Ensemble Learning for BraTS 2025 Pediatric Brain Tumor Segmentation

Yuxiao Yi, Qingyao Zhuang, Zhi-Qin John Xu, Xiaowen Wang, Yan Ren, Tianming Qiu

TL;DR

This work tackles pediatric brain tumor segmentation by proposing a frequency-aware ensemble of nnU-Net, Swin UNETR, and HFF-Net for the BraTS-PED 2025 challenge. It introduces three extensions: tunable initialization for nnU-Net, transfer learning from BraTS 2021 to improve Swin UNETR generalization, and a frequency-domain decomposition in HFF-Net to separate low-frequency tissue contours from high-frequency texture details. Through five-fold cross-validation and a carefully designed data preprocessing pipeline (skull stripping and LF/HF decompositions), the ensemble achieves competitive Dice scores across six subregions and ranks 1st on the BraTS-PED 2025 test set, demonstrating robustness on unseen pediatric data. The approach outputs clinically meaningful segmentation accuracy while paving the way for broader applications in pediatric neuro-oncology through enhanced multi-model fusion and frequency-aware processing.

Abstract

Pediatric brain tumor segmentation presents unique challenges due to the rarity and heterogeneity of these malignancies, yet remains critical for clinical diagnosis and treatment planning. We propose an ensemble approach integrating nnU-Net, Swin UNETR, and HFF-Net for the BraTS-PED 2025 challenge. Our method incorporates three key extensions: adjustable initialization scales for optimal nnU-Net complexity control, transfer learning from BraTS 2021 pre-trained models to enhance Swin UNETR's generalization on pediatric dataset, and frequency domain decomposition for HFF-Net to separate low-frequency tissue contours from high-frequency texture details. Our final ensemble framework combines nnU-Net ($γ=0.7$), fine-tuned Swin UNETR, and HFF-Net, achieving Dice scores of 62.7% (CC), 83.2% (ED), 72.9% (ET), 85.7% (NET), 91.8% (TC), and 92.6% (WT) on the unseen test dataset, respectively. Our proposed method achieves first place (rank 1st) in the BraTS 2025 Pediatric Brain Tumor Segmentation Challenge.

Frequency-Aware Ensemble Learning for BraTS 2025 Pediatric Brain Tumor Segmentation

TL;DR

This work tackles pediatric brain tumor segmentation by proposing a frequency-aware ensemble of nnU-Net, Swin UNETR, and HFF-Net for the BraTS-PED 2025 challenge. It introduces three extensions: tunable initialization for nnU-Net, transfer learning from BraTS 2021 to improve Swin UNETR generalization, and a frequency-domain decomposition in HFF-Net to separate low-frequency tissue contours from high-frequency texture details. Through five-fold cross-validation and a carefully designed data preprocessing pipeline (skull stripping and LF/HF decompositions), the ensemble achieves competitive Dice scores across six subregions and ranks 1st on the BraTS-PED 2025 test set, demonstrating robustness on unseen pediatric data. The approach outputs clinically meaningful segmentation accuracy while paving the way for broader applications in pediatric neuro-oncology through enhanced multi-model fusion and frequency-aware processing.

Abstract

Pediatric brain tumor segmentation presents unique challenges due to the rarity and heterogeneity of these malignancies, yet remains critical for clinical diagnosis and treatment planning. We propose an ensemble approach integrating nnU-Net, Swin UNETR, and HFF-Net for the BraTS-PED 2025 challenge. Our method incorporates three key extensions: adjustable initialization scales for optimal nnU-Net complexity control, transfer learning from BraTS 2021 pre-trained models to enhance Swin UNETR's generalization on pediatric dataset, and frequency domain decomposition for HFF-Net to separate low-frequency tissue contours from high-frequency texture details. Our final ensemble framework combines nnU-Net (), fine-tuned Swin UNETR, and HFF-Net, achieving Dice scores of 62.7% (CC), 83.2% (ED), 72.9% (ET), 85.7% (NET), 91.8% (TC), and 92.6% (WT) on the unseen test dataset, respectively. Our proposed method achieves first place (rank 1st) in the BraTS 2025 Pediatric Brain Tumor Segmentation Challenge.

Paper Structure

This paper contains 14 sections, 1 equation, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Data Description
  4. Data Pre-processing
  5. Model Training
  6. nnU-Net
  7. Swin UNETR
  8. HFF-Net
  9. Model Ensemble
  10. Evaluation Metrics
  11. Results
  12. Quantitative results
  13. Qualitative Results
  14. Discussion

Figures (3)

  • Figure 1: Frequency decomposition results of T2-FLAIR on the training set. LF denotes low frequency component and HF1$\sim$HF4 represents four high frequency components, respectively.
  • Figure 2: Frequency decomposition results of T1C on the training set. LF denotes low frequency component and HF1$\sim$HF4 represents four high frequency components, respectively.
  • Figure 3: Quantitative results of the final model on the validation set. The three selected examples correspond to BraTS-PED-00310-000, BraTS-PED-00315-000, and BraTS-PED-00318-000, respectively.