Scholarlink

<p>Accurately segmenting tumors from 3D medical images is crucial for various clinical applications. While deep networks based on local convolutions stand out as the preferred choice for various segmentation tasks, transformers, which process inputs as 1D sequences, have gained attention as a powerful alternative due to their proficiency in capturing long-range spatial contexts. Despite their success, standard self-attention-based transformers still face challenges in resource-constrained environments because of their quadratic computational complexity and reliance on large-scale pre-training. In this work, we present MM-Net, a convolution-transformer hybrid architecture that can synergize local convolution with self-attention mechanisms. This approach allows for the capture of multi-scale, multi-view contextual information, which is essential for accurately delineating complexly shaped objects across different scales. To demonstrate the efficacy of the proposed lightweight model, we conduct comprehensive evaluations and ablation studies on three brain tumor datasets alongside a liver tumor dataset. The results show that MM-Net achieves comparable or better performance over state-of-the-art methodologies. Additionally, the proposed model has significantly fewer learnable parameters and reduced floating point operations (FLOPs), making it particularly beneficial for clinical 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

MM-Net: accurate tumor segmentation from medical images with lightweight hybrid transformers

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Peixu Wang,
  • Baijia Chen,
  • Shikun Liu,
  • Ye Wang,
  • Jialin Peng

摘要

Accurately segmenting tumors from 3D medical images is crucial for various clinical applications. While deep networks based on local convolutions stand out as the preferred choice for various segmentation tasks, transformers, which process inputs as 1D sequences, have gained attention as a powerful alternative due to their proficiency in capturing long-range spatial contexts. Despite their success, standard self-attention-based transformers still face challenges in resource-constrained environments because of their quadratic computational complexity and reliance on large-scale pre-training. In this work, we present MM-Net, a convolution-transformer hybrid architecture that can synergize local convolution with self-attention mechanisms. This approach allows for the capture of multi-scale, multi-view contextual information, which is essential for accurately delineating complexly shaped objects across different scales. To demonstrate the efficacy of the proposed lightweight model, we conduct comprehensive evaluations and ablation studies on three brain tumor datasets alongside a liver tumor dataset. The results show that MM-Net achieves comparable or better performance over state-of-the-art methodologies. Additionally, the proposed model has significantly fewer learnable parameters and reduced floating point operations (FLOPs), making it particularly beneficial for clinical applications.