摘要
卷积神经网络(CNN)在医学图像分割中取得了显著的进展,但其在捕捉长距离依赖信息方面存在局限性.虽然Transformer模型在处理远程依赖方面表现出色,但自注意力机制导致了较高的计算成本.为了解决这些问题,提出了多尺度医学图像分割网络VMA-UNet(VMamba ASPP U-Net),它融合了VMamba的VSS块结构和空洞空间卷积池化金字塔(ASPP)模块.VMA-UNet利用VSS块的线性复杂度特性,实现高效的全局信息建模,并结合ASPP模块在多个尺度上捕捉医学图像中的关键特征.通过在ACDC、COVID-19 CT和Synapse等数据集上的广泛实验,结果表明:VMA-UNet在分割精度和计算效率上均优于基于CNN和Transformer方法,显示了其在不同任务中的竞争力.VMA-UNet克服了CNN在捕捉远程信息方面的局限性,实现了高效的多尺度建模,展现了其在医学图像分割中的巨大潜力.
Convolutional Neural Networks(CNN)have achieved remarkable progress in medical image segmentation,but it exhibits limitations in capturing long-range dependencies.While Transformer models excel in handling long-range dependencies,the self-attention mechanism incurs high computational costs.To address these issues,VMA-UNet(VMamba ASPP U-Net)is proposed,it is a multi-scale medical image segmentation network that integrates the VSS(Visual State-Space)block structure from VMamba and the Atrous Spatial Pyramid Pooling(ASPP)module.VMA-UNet leverages the linear complexity of the VSS block to enable efficient global information modeling and incorporates the ASPP module to capture critical features of medical images at multiple scales.Extensive experiments conducted on datasets including ACDC,COVID-19 CT,and Synapse demonstrate that VMA-UNet outperforms CNN-and Transformer-based methods in terms of segmentation accuracy and computational efficiency.By overcoming the limitations of CNN in modeling long-range dependencies and enabling efficient multi-scale modeling,VMA-UNet showcases its immense potential in medical image segmentation tasks.
作者
王海
李亚鸽
林愉萱
陆雪松
WANG Hai;LI Yage;LIN Yuxuan;LU Xuesong(College of Biomedical Engineering,South-Central Minzu University,Wuhan 430074,China)
出处
《中南民族大学学报(自然科学版)》
2026年第1期51-59,共9页
Journal of South-Central Minzu University(Natural Science Edition)
基金
湖北省自然科学基金资助项目(2016CFB489)
中央高校基本科研业务费专项资助项目(CZZ24014)。
