Scholarlink

<p>This study presents SEARNet (Segmentation Network Enhanced with Attention and Residual Learning Network), a novel deep learning architecture specifically designed to address the challenges of breast ultrasound (BUS) lesion segmentation, such as low contrast, speckle noise, and irregular lesion boundaries. SEARNet introduces three key innovations, namely (i) Squeeze and Excitation (SE)-assisted residual convolutional blocks to enhance feature recalibration and gradient flow, (ii) an attention-based gating mechanism in skip connections to selectively transfer relevant lesion features, and (iii) dilated convolutions in the bottleneck layer to enlarge the receptive field without increasing parameters, enabling the network to preserve spatial details while capturing broader contextual information. These architectural contributions collectively improve the ability to delineate lesion boundaries accurately, even under challenging imaging conditions. Extensive experiments were conducted on three publicly available datasets, BUSI, BUS-BRA, and a brain tumor MRI dataset, to evaluate robustness and generalization across different imaging modalities. On the BUSI dataset, SEARNet achieved an F1-score of 85.31% and a Jaccard (IoU) index of 78.07%, surpassing state-of-the-art segmentation models including U-Net, DeepLabV3+, and TransUNet. Moreover, SEARNet maintains a favorable balance between segmentation accuracy and computational efficiency, with 25.7&#xa0;million parameters and 29.4 GFLOPs, making it suitable for clinical deployment in resource-constrained environments. The integration of residual learning, channel attention, and attention-guided skip connections enables SEARNet to deliver superior lesion boundary delineation, demonstrating strong potential for real-world breast cancer detection and other diagnostic imaging applications.</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

Attention-guided residual learning with SEARNet for improved breast ultrasound lesion segmentation

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Berkay Eren

摘要

This study presents SEARNet (Segmentation Network Enhanced with Attention and Residual Learning Network), a novel deep learning architecture specifically designed to address the challenges of breast ultrasound (BUS) lesion segmentation, such as low contrast, speckle noise, and irregular lesion boundaries. SEARNet introduces three key innovations, namely (i) Squeeze and Excitation (SE)-assisted residual convolutional blocks to enhance feature recalibration and gradient flow, (ii) an attention-based gating mechanism in skip connections to selectively transfer relevant lesion features, and (iii) dilated convolutions in the bottleneck layer to enlarge the receptive field without increasing parameters, enabling the network to preserve spatial details while capturing broader contextual information. These architectural contributions collectively improve the ability to delineate lesion boundaries accurately, even under challenging imaging conditions. Extensive experiments were conducted on three publicly available datasets, BUSI, BUS-BRA, and a brain tumor MRI dataset, to evaluate robustness and generalization across different imaging modalities. On the BUSI dataset, SEARNet achieved an F1-score of 85.31% and a Jaccard (IoU) index of 78.07%, surpassing state-of-the-art segmentation models including U-Net, DeepLabV3+, and TransUNet. Moreover, SEARNet maintains a favorable balance between segmentation accuracy and computational efficiency, with 25.7 million parameters and 29.4 GFLOPs, making it suitable for clinical deployment in resource-constrained environments. The integration of residual learning, channel attention, and attention-guided skip connections enables SEARNet to deliver superior lesion boundary delineation, demonstrating strong potential for real-world breast cancer detection and other diagnostic imaging applications.