Scholarlink

<p>3D visualization of hippocampus and amygdala plays a vital role in pre-operative stereotactic surgery planning. The hippocampus and amygdala, critical brain structures associated with memory and emotion, present significant challenges for accurate and simultaneous segmentation due to their close anatomical proximity. This study introduces a Mamba-enhanced hierarchical attention network (MEHA-Net) specifically developed for precise concurrent segmentation of these structures in brain MRI images. The Mamba model is an advanced state-space module. It is integrated as a residual structure within the convolutional network backbone, improving local feature extraction while maintaining low model complexity. Additionally, a hierarchical attention mechanism combines low-level attention gates to suppress irrelevant regions with high-level self-attention from transformers to capture global context and long-range dependencies. Evaluated on a dataset of 100 brain MRI images, the proposed method achieved Dice coefficients of 93.4% for hippocampus segmentation and 91.0% for amygdala segmentation, significantly outperforming state-of-the-art models. These results highlight the efficacy of MEHA-Net in dense data scenarios, demonstrating its superiority in simultaneous segmentation of the hippocampus and amygdala. Code is available at: <a href="https://github.com/chrisma98/MEHANet">https://github.com/chrisma98/MEHANet</a>.</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

Mamba-enhanced hierarchical attention network for precise visualization of hippocampus and amygdala

  • 期刊论文

  • 发表日期：

    2025年2月3日

查看全文
  • Junchi Ma,
  • Yuanqing Wang,
  • Guangmiao Ding,
  • Wei Cao,
  • Xiangyun Liao,
  • Ping Zhang,
  • Jianping Lv

摘要

3D visualization of hippocampus and amygdala plays a vital role in pre-operative stereotactic surgery planning. The hippocampus and amygdala, critical brain structures associated with memory and emotion, present significant challenges for accurate and simultaneous segmentation due to their close anatomical proximity. This study introduces a Mamba-enhanced hierarchical attention network (MEHA-Net) specifically developed for precise concurrent segmentation of these structures in brain MRI images. The Mamba model is an advanced state-space module. It is integrated as a residual structure within the convolutional network backbone, improving local feature extraction while maintaining low model complexity. Additionally, a hierarchical attention mechanism combines low-level attention gates to suppress irrelevant regions with high-level self-attention from transformers to capture global context and long-range dependencies. Evaluated on a dataset of 100 brain MRI images, the proposed method achieved Dice coefficients of 93.4% for hippocampus segmentation and 91.0% for amygdala segmentation, significantly outperforming state-of-the-art models. These results highlight the efficacy of MEHA-Net in dense data scenarios, demonstrating its superiority in simultaneous segmentation of the hippocampus and amygdala. Code is available at: https://github.com/chrisma98/MEHANet.