Scholarlink

<p>Breast cancer is a major cause of cancer-related deaths among women. Ultrasound imaging is commonly used for early detection due to its non-invasive and cost-effective nature. However, accurately segmenting breast tumors in ultrasound images is difficult due to speckle noise, irregular tumor boundaries, and low contrast between tissues. To address these challenges, a U-Net-based architecture is proposed, which incorporates multi-resolution analysis by performing discrete wavelet transform (DWT) and max pooling operations in parallel at each encoder stage. Unlike prior wavelet-convolutional neural network (CNN) hybrid methods, which discard one or more of the four subbands —low–low (LL), low–high (LH), high–low (HL), and high–high (HH) —the proposed architecture preserves all subbands. These subbands are combined with spatial features extracted through max pooling. By extracting both spatial and frequency-domain features in parallel and concatenating them, the model captures broad structural patterns as well as fine directional details. This leads to a richer and more discriminative feature representation, enhancing tumor segmentation performance. The encoder also uses multi-step downsampling short connections (MDSC) to enhance feature flow and support cross-scale fusion, preserving both context and boundary details. Additionally, multi-out U-Net (MOU) blocks replace traditional skip connections, enhancing feature bridging between the encoder and decoder by incorporating multi-resolution frequency cues. These architectural innovations collectively enhance segmentation performance on challenging breast ultrasound datasets, outperforming current state-of-the-art methods.</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 novel approach for breast tumor segmentation using multi-resolution analysis through wavelet transform

  • 期刊论文

  • 发表日期：

    2026年2月16日

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

摘要

Breast cancer is a major cause of cancer-related deaths among women. Ultrasound imaging is commonly used for early detection due to its non-invasive and cost-effective nature. However, accurately segmenting breast tumors in ultrasound images is difficult due to speckle noise, irregular tumor boundaries, and low contrast between tissues. To address these challenges, a U-Net-based architecture is proposed, which incorporates multi-resolution analysis by performing discrete wavelet transform (DWT) and max pooling operations in parallel at each encoder stage. Unlike prior wavelet-convolutional neural network (CNN) hybrid methods, which discard one or more of the four subbands —low–low (LL), low–high (LH), high–low (HL), and high–high (HH) —the proposed architecture preserves all subbands. These subbands are combined with spatial features extracted through max pooling. By extracting both spatial and frequency-domain features in parallel and concatenating them, the model captures broad structural patterns as well as fine directional details. This leads to a richer and more discriminative feature representation, enhancing tumor segmentation performance. The encoder also uses multi-step downsampling short connections (MDSC) to enhance feature flow and support cross-scale fusion, preserving both context and boundary details. Additionally, multi-out U-Net (MOU) blocks replace traditional skip connections, enhancing feature bridging between the encoder and decoder by incorporating multi-resolution frequency cues. These architectural innovations collectively enhance segmentation performance on challenging breast ultrasound datasets, outperforming current state-of-the-art methods.