Scholarlink

<p>Breast cancer is one of the leading causes of cancer-related deaths among women worldwide. Accurate tumor segmentation is crucial for early detection, effective treatment, and better survival rates. However, low contrast, noise, and intensity variations in breast imaging pose major segmentation issues. Existing approaches based on the spatial domain often fail to address these limitations, thereby limiting segmentation accuracy. The UNetDSE-DCT model, a UNet-based model with deeper layers that combine features of spatial and transform domains, is proposed to overcome these challenges. It combines features of max-pooling for spatial and discrete cosine transform pooling for the transform domain at each encoder layer, allowing the model to capture complex morphological forms and improve border delineation. In addition, standard skip connections are replaced with squeeze-and-excitation blocks to highlight key features while suppressing irrelevant ones. Furthermore, the guided loss mechanism is applied at each decoder layer to address class imbalance. The model was evaluated on three publicly available datasets: BUSIS, Dataset B, and BUSI. Our findings demonstrate that UNetDSE-DCT is an efficient and accurate segmentation model that outperforms the state-of-the-art methods in performance, space, and time complexity.</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

A Transform-Based UNet Architecture for Breast Tumor Segmentation

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Chandan Singh,
  • Sukhjeet Kaur,
  • Heena Jasrotia

摘要

Breast cancer is one of the leading causes of cancer-related deaths among women worldwide. Accurate tumor segmentation is crucial for early detection, effective treatment, and better survival rates. However, low contrast, noise, and intensity variations in breast imaging pose major segmentation issues. Existing approaches based on the spatial domain often fail to address these limitations, thereby limiting segmentation accuracy. The UNetDSE-DCT model, a UNet-based model with deeper layers that combine features of spatial and transform domains, is proposed to overcome these challenges. It combines features of max-pooling for spatial and discrete cosine transform pooling for the transform domain at each encoder layer, allowing the model to capture complex morphological forms and improve border delineation. In addition, standard skip connections are replaced with squeeze-and-excitation blocks to highlight key features while suppressing irrelevant ones. Furthermore, the guided loss mechanism is applied at each decoder layer to address class imbalance. The model was evaluated on three publicly available datasets: BUSIS, Dataset B, and BUSI. Our findings demonstrate that UNetDSE-DCT is an efficient and accurate segmentation model that outperforms the state-of-the-art methods in performance, space, and time complexity.