Scholarlink

In this study, we explore the effectiveness of a stacking ensemble approach for brain tumor segmentation by combining three distinct deep neural network architectures: nnU-Net (based on convolutional neural network, SwinUNETR (based on transformer), and LoG-VMamba (based on VMamba). Each model offers unique representational strengths that, when integrated through ensemble learning, enhance overall segmentation performance. To enable robust fusion, our model utilizes two core components: (1) label noise modeling through signed soft labels and (2) stacking-based model fusion. All models, including ensembles, are evaluated on two tasks of the BraTS 2025 challenge: Task 1 (pre- and post-treatment glioma in adults) and Task 5 (glioma segmentation within a Sub-Saharan African patient population). Our results show that SwinUNETR consistently performs below nnU-Net and LoG-VMamba on both tasks. The ensemble method surpasses all individual models according to both regional and boundary-aware metrics, demonstrating improved accuracy and robustness. Our findings underscore the value of architectural diversity and ensemble learning in advancing medical image segmentation.

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

Ensembling CNN, Transformer, and Mamba with Stacking for Brain Tumor Segmentation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Trung D. Q. Dang,
  • Huy Hoang Nguyen,
  • Aleksei Tiulpin

摘要

In this study, we explore the effectiveness of a stacking ensemble approach for brain tumor segmentation by combining three distinct deep neural network architectures: nnU-Net (based on convolutional neural network, SwinUNETR (based on transformer), and LoG-VMamba (based on VMamba). Each model offers unique representational strengths that, when integrated through ensemble learning, enhance overall segmentation performance. To enable robust fusion, our model utilizes two core components: (1) label noise modeling through signed soft labels and (2) stacking-based model fusion. All models, including ensembles, are evaluated on two tasks of the BraTS 2025 challenge: Task 1 (pre- and post-treatment glioma in adults) and Task 5 (glioma segmentation within a Sub-Saharan African patient population). Our results show that SwinUNETR consistently performs below nnU-Net and LoG-VMamba on both tasks. The ensemble method surpasses all individual models according to both regional and boundary-aware metrics, demonstrating improved accuracy and robustness. Our findings underscore the value of architectural diversity and ensemble learning in advancing medical image segmentation.