Scholarlink

<p>Medical image segmentation is constrained by the computational resources of mobile medical devices, making computation-intensive models unsuitable. State space models, represented by Mamba, can effectively model long-range interactions while maintaining linear computational complexity. However, current lightweight architectures based on Mamba struggle to balance segmentation accuracy. To address this, we propose SSRepVM-UNet. Specifically, we introduce the SS-RepVM module, which leverages the lightweight convolutional RepViT module and the SS2D module with a scan expansion strategy to process different channels of feature maps in parallel. Our network significantly reduces the number of parameters and computational load while ensuring excellent performance. We conducted extensive experiments on four public dermoscopy lesion datasets and two polyp segmentation datasets. The results demonstrate that our network is highly competitive in medical image segmentation tasks, achieving state-of-the-art (SOTA) performance on the ISIC2018 dataset with only 0.17M parameters and 0.18 GFLOPs, setting a new benchmark for lightweight models.</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

SSRepVM-UNet: a lightweight hybrid model for medical image segmentation based on channel parallelism

  • 期刊论文

  • 发表日期：

    2025年8月8日

查看全文
  • Yijing Guo,
  • Fuhang Li,
  • Kunhua Li,
  • Huawei Wang,
  • Pengyu Xu

摘要

Medical image segmentation is constrained by the computational resources of mobile medical devices, making computation-intensive models unsuitable. State space models, represented by Mamba, can effectively model long-range interactions while maintaining linear computational complexity. However, current lightweight architectures based on Mamba struggle to balance segmentation accuracy. To address this, we propose SSRepVM-UNet. Specifically, we introduce the SS-RepVM module, which leverages the lightweight convolutional RepViT module and the SS2D module with a scan expansion strategy to process different channels of feature maps in parallel. Our network significantly reduces the number of parameters and computational load while ensuring excellent performance. We conducted extensive experiments on four public dermoscopy lesion datasets and two polyp segmentation datasets. The results demonstrate that our network is highly competitive in medical image segmentation tasks, achieving state-of-the-art (SOTA) performance on the ISIC2018 dataset with only 0.17M parameters and 0.18 GFLOPs, setting a new benchmark for lightweight models.