Scholarlink

The 2025 BraTS Generalizability Across Tumors (GoAT) challenge highlights the need for segmentation models that perform robustly across diverse brain tumor types. We propose a segmentation pipeline that integrates pseudo-label supervised fine-tuning, architectural ensembling, and a novel ratio-adaptive postprocessing strategy. By generating high-quality pseudo-labels from unlabeled cases and fine-tuning multiple nnUNet variants, we enhance model generalization, particularly for whole tumor segmentation. Our ensemble of heterogeneous architectures—including ResNetEncoder variants and U-MambaBot—further improves performance across all tumor subregions. To refine predictions, we introduce a ratio-adaptive thresholding scheme that dynamically adjusts postprocessing cutoffs based on predicted tumor volume, achieving a better balance between precision and recall. Together, these strategies yield strong lesion-wise Dice scores across tumor compartments on the validation set, demonstrating effective tumor-type-agnostic segmentation. Our final test-phase submission achieved lesion-wise Dice scores of 0.828 (WT), 0.827 (TC), and 0.800 (ET), co-ranking first place in the BraTS 2025 Lighthouse GoAT Challenge. We also explore a multitask anatomical segmentation approach as a potential future direction.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing Brain Tumor Segmentation Generalizability via Pseudo-Labeling and Ratio-Adaptive Postprocessing

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • To-Liang Hsu,
  • Dang Khoa Nguyen,
  • Pai Lin,
  • Ching-Ting Lin,
  • Wei-Chun Wang

摘要

The 2025 BraTS Generalizability Across Tumors (GoAT) challenge highlights the need for segmentation models that perform robustly across diverse brain tumor types. We propose a segmentation pipeline that integrates pseudo-label supervised fine-tuning, architectural ensembling, and a novel ratio-adaptive postprocessing strategy. By generating high-quality pseudo-labels from unlabeled cases and fine-tuning multiple nnUNet variants, we enhance model generalization, particularly for whole tumor segmentation. Our ensemble of heterogeneous architectures—including ResNetEncoder variants and U-MambaBot—further improves performance across all tumor subregions. To refine predictions, we introduce a ratio-adaptive thresholding scheme that dynamically adjusts postprocessing cutoffs based on predicted tumor volume, achieving a better balance between precision and recall. Together, these strategies yield strong lesion-wise Dice scores across tumor compartments on the validation set, demonstrating effective tumor-type-agnostic segmentation. Our final test-phase submission achieved lesion-wise Dice scores of 0.828 (WT), 0.827 (TC), and 0.800 (ET), co-ranking first place in the BraTS 2025 Lighthouse GoAT Challenge. We also explore a multitask anatomical segmentation approach as a potential future direction.