Scholarlink

<p>Accurate 3D brain-tumor segmentation has been recognized as a critical step for diagnosis, treatment planning, and outcome prediction in neuro-oncology. Although numerous 3D deep learning architectures have been developed, existing transformer-based models have suffered from excessive computational cost and weak boundary representation. To address these challenges, an EESU-Net: Edge-Enhanced Swin-Transformer Network has been proposed for accurate and efficient volumetric brain-tumor segmentation. The architecture has integrated a Swin-Transformer encoder for global context extraction, an Edge-Enhancement Module (EEM) employing Sobel and Laplacian filters to strengthen tumor boundary localization, and a Lightweight Convolutional Attention (LCA) decoder for fine-grained feature reconstruction. A hybrid dice–boundary loss has been utilized to jointly optimize region overlap and contour precision. Experimental results on the BraTS2021 benchmark have been demonstrated. The EESU-Net has achieved a mean Dice of 96.2 ± 0.3% and HD95 of 3.8 ± 0.2&#xa0;mm, outperforming recent state-of-the-art 3D architectures, while using only approximately 19.6&#xa0;million parameters. These findings have confirmed that EESU-Net has achieved a superior trade-off between accuracy, computational efficiency, and boundary fidelity, establishing it as a clinically scalable solution for 3D brain-tumor segmentation.</p>

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

EESU-Net: edge-enhanced swin-transformer network for accurate and efficient 3D brain-tumor segmentation

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Nattavut Sriwiboon

摘要

Accurate 3D brain-tumor segmentation has been recognized as a critical step for diagnosis, treatment planning, and outcome prediction in neuro-oncology. Although numerous 3D deep learning architectures have been developed, existing transformer-based models have suffered from excessive computational cost and weak boundary representation. To address these challenges, an EESU-Net: Edge-Enhanced Swin-Transformer Network has been proposed for accurate and efficient volumetric brain-tumor segmentation. The architecture has integrated a Swin-Transformer encoder for global context extraction, an Edge-Enhancement Module (EEM) employing Sobel and Laplacian filters to strengthen tumor boundary localization, and a Lightweight Convolutional Attention (LCA) decoder for fine-grained feature reconstruction. A hybrid dice–boundary loss has been utilized to jointly optimize region overlap and contour precision. Experimental results on the BraTS2021 benchmark have been demonstrated. The EESU-Net has achieved a mean Dice of 96.2 ± 0.3% and HD95 of 3.8 ± 0.2 mm, outperforming recent state-of-the-art 3D architectures, while using only approximately 19.6 million parameters. These findings have confirmed that EESU-Net has achieved a superior trade-off between accuracy, computational efficiency, and boundary fidelity, establishing it as a clinically scalable solution for 3D brain-tumor segmentation.