We propose an ensemble approach integrating nnU-Net, Swin UNETR, and HFFNet 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 model to enhance Swin UNETR’s generalization on pediatric dataset.
- Frequency domain decomposition for HFF-Net to separate low-frequency tissue contours from high-frequency texture details.
🚩[2025.12] Our team received an officially issued electronic certificate.
🚩[2025.11] The source code of our solution has been released.
🚩[2025.10] Our solution achieves 🥇rank 1st in the BraTS 2025 Pediatric Brain Tumor Segmentation Challenge.
🚩[2025.09] We are invited to give an oral presentation during the MICCAI BraTS 2025 Challenge Workshop on 23 September 2025.
Our work builds upon several excellent prior methods. We thank the authors for open-sourcing their code, including the famous nnUnet, Swin-UNETR, and the recent HFF-Net. For more implementation details, please refer to their original repositories.