Scholarlink

To address the challenge of generalizing brain tumor segmentation across diverse tumor types, the BraTS 2025 Challenge introduced the Generalizability Across Tumors (GoAT) task, focusing on robustness under distribution shifts. In response, we propose ADMFNet (Adaptively Modulated DMF Network), a novel 3D segmentation network designed to enhance cross-distribution performance in multimodal MRI. ADMFNet features a multi-scale encoder-decoder architecture based on dilated convolutional units (DMFUnit) and incorporates a lightweight Adaptation Module to reduce feature mismatch between encoder and decoder stages, improving cross-modal fusion. The model is trained on the BraTS-GoAT dataset, which includes four MRI modalities: T1n, T1c, T2f, and T2w. Ground truth follows BraTS annotation standards, and training uses class-balanced Generalized Dice Loss. On the GoAT testing set, ADMFNet achieves Dice scores of 0.696 (WT), 0.744 (TC), and 0.717 (ET), with corresponding NSD (1.0 mm) values of 0.616, 0.708, and 0.732. The model shows excellent performance on clearly defined regions such as edema and radiation cavities (Dice = 1.000), and maintains solid results on challenging areas including non-enhancing tumor cores and cystic components.

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

ADMFNet: Enhancing Cross-Tumor Generalization in Multi-Modal MRI Segmentation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hongjuan Wang,
  • Yixin Zhang,
  • Jindong Sun,
  • Xinjun An,
  • Liying Zhu,
  • Chunyao Li

摘要

To address the challenge of generalizing brain tumor segmentation across diverse tumor types, the BraTS 2025 Challenge introduced the Generalizability Across Tumors (GoAT) task, focusing on robustness under distribution shifts. In response, we propose ADMFNet (Adaptively Modulated DMF Network), a novel 3D segmentation network designed to enhance cross-distribution performance in multimodal MRI. ADMFNet features a multi-scale encoder-decoder architecture based on dilated convolutional units (DMFUnit) and incorporates a lightweight Adaptation Module to reduce feature mismatch between encoder and decoder stages, improving cross-modal fusion. The model is trained on the BraTS-GoAT dataset, which includes four MRI modalities: T1n, T1c, T2f, and T2w. Ground truth follows BraTS annotation standards, and training uses class-balanced Generalized Dice Loss. On the GoAT testing set, ADMFNet achieves Dice scores of 0.696 (WT), 0.744 (TC), and 0.717 (ET), with corresponding NSD (1.0 mm) values of 0.616, 0.708, and 0.732. The model shows excellent performance on clearly defined regions such as edema and radiation cavities (Dice = 1.000), and maintains solid results on challenging areas including non-enhancing tumor cores and cystic components.