Scholarlink

<p>Accurate echocardiographic segmentation, crucial for cardiac assessment, remains challenging due to poor image quality, significant anatomical variability, and the inherent difficulty in simultaneously capturing fine local details and broad global context. While Mamba, a state-of-the-art state-space model, enables efficient global context modeling with linear complexity, its reliance on a unidirectional selective scan inherently limits its ability to model the non-sequential structure of 2D image data. To address this limitation, we propose a novel segmentation framework EchoUMamba, which leverages an advanced EchoVSSBlock to better integrate the rich information of features extracted from four directions. Experiments on the EchoNet-Dynamic and CAMUS datasets show that EchoUMamba achieves competitive segmentation performance with Dice = 0.9336 and IoU = 0.8934, while maintaining computational efficiency with 26.3M parameters and 32.7G FLOPs.</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

Efficient Echocardiogram Segmentation with Multi-Scale Residual State Space Modeling

  • 期刊论文

  • 发表日期：

    2025年10月28日

查看全文
  • Zihe Luo,
  • Fufeng Wang,
  • Xinkui Liao,
  • Yuheng Liang,
  • Kaihua Che,
  • Wei Lv,
  • Xiaolin Zhu

摘要

Accurate echocardiographic segmentation, crucial for cardiac assessment, remains challenging due to poor image quality, significant anatomical variability, and the inherent difficulty in simultaneously capturing fine local details and broad global context. While Mamba, a state-of-the-art state-space model, enables efficient global context modeling with linear complexity, its reliance on a unidirectional selective scan inherently limits its ability to model the non-sequential structure of 2D image data. To address this limitation, we propose a novel segmentation framework EchoUMamba, which leverages an advanced EchoVSSBlock to better integrate the rich information of features extracted from four directions. Experiments on the EchoNet-Dynamic and CAMUS datasets show that EchoUMamba achieves competitive segmentation performance with Dice = 0.9336 and IoU = 0.8934, while maintaining computational efficiency with 26.3M parameters and 32.7G FLOPs.