Scholarlink

Purpose <p>Brain segmentation is crucial for analyzing and understanding complex neurological structures and functions, enabling precise diagnosis, treatment planning, and research in various medical and scientific fields. Current brain segmentation methods face challenges in accurately outlining complex structures, dealing with imaging artifacts, and adapting to individual and modality differences, which may impact accuracy and applicability.</p> Methods <p>We propose a novel deep learning model, 3D SE-Res-UNet, for brain tumor segmentation in multi-parametric magnetic resonance imaging (mpMRI) scans. This architecture incorporates Squeeze-and-Excitation (SE) layers to enhance feature re-calibration and Residual (Res) layers for efficient information flow. The encoder utilizes SE layers to capture complex hierarchical features, while the decoder employs Res layers for precise reconstruction. Batch normalization and Leaky ReLU activation functions maintain stability and non-linearity throughout the network.</p> Results <p>The proposed 3D SE-Res-UNet model demonstrated high performance in medical image segmentation on the BraTS’21 dataset, achieving substantial overlap between predicted and ground truth masks as evidenced by Dice Similarity Coefficient (DSC) values of 0.8968 for Tumor Core (TC), 0.9342 for Whole Tumor (WT), and 0.8621 for Enhancing Tumor (ET).</p> Conclusion <p>The accuracy validated by favorable Hausdorff Distance (HD-95), Mean Intersection over Union (mIoU), Surface Distance metrics, and Surface Dice coefficients, underscores its effectiveness in capturing proximity and geometric agreement between predicted and true tumor surfaces. The proposed model had achieved competitive results with very low computational complexity against state-of-the-art models.</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

3D SE-Res-UNet: squeeze-excitation enhanced residual UNet for brain tumor segmentation in multi-modal MRI scans

  • 期刊论文

  • 发表日期：

    2025年8月4日

查看全文
  • Praveen M A,
  • Srilatha Chebrolu

摘要

Purpose

Brain segmentation is crucial for analyzing and understanding complex neurological structures and functions, enabling precise diagnosis, treatment planning, and research in various medical and scientific fields. Current brain segmentation methods face challenges in accurately outlining complex structures, dealing with imaging artifacts, and adapting to individual and modality differences, which may impact accuracy and applicability.

Methods

We propose a novel deep learning model, 3D SE-Res-UNet, for brain tumor segmentation in multi-parametric magnetic resonance imaging (mpMRI) scans. This architecture incorporates Squeeze-and-Excitation (SE) layers to enhance feature re-calibration and Residual (Res) layers for efficient information flow. The encoder utilizes SE layers to capture complex hierarchical features, while the decoder employs Res layers for precise reconstruction. Batch normalization and Leaky ReLU activation functions maintain stability and non-linearity throughout the network.

Results

The proposed 3D SE-Res-UNet model demonstrated high performance in medical image segmentation on the BraTS’21 dataset, achieving substantial overlap between predicted and ground truth masks as evidenced by Dice Similarity Coefficient (DSC) values of 0.8968 for Tumor Core (TC), 0.9342 for Whole Tumor (WT), and 0.8621 for Enhancing Tumor (ET).

Conclusion

The accuracy validated by favorable Hausdorff Distance (HD-95), Mean Intersection over Union (mIoU), Surface Distance metrics, and Surface Dice coefficients, underscores its effectiveness in capturing proximity and geometric agreement between predicted and true tumor surfaces. The proposed model had achieved competitive results with very low computational complexity against state-of-the-art models.