基于Mamba模型的U型医学图像分割网络

U-Shaped Medical Image Segmentation Network Based on the Mamba Model

摘要: 在医学图像分割网络结构中，卷积神经网络(CNN)无疑是最强大的网络，它凭借强大的局部特征提取能力，取得了很多突破性进展，成为该领域的主流方法之一。然而，CNN固有的局部感知机制使其在捕捉图像中长距离依赖信息时存在明显的局限性，这在一定程度上制约了其分割性能。Transformer模型凭借自注意力机制，在处理长距离依赖关系上展现出显著优势，能够有效挖掘图像全局范围内的特征关系。但自注意力机制的计算复杂度随输入序列长度的平方而增加，导致Transformer模型在处理高分辨率医学影像时，往往面临计算成本过高、内存占用量大的问题。为解决该问题，本文提出一种新的分割网络——M-UNet。该网络的创新在于将Mamba模型特有的VSS块结构与空洞空间卷积金字塔池化模块进行结合。利用VSS块的线性复杂度特性，在实现高效全局信息建模的同时，大幅降低计算内存。其次，借助ASPP模块多尺度空洞卷积的优势，精准捕捉医学图像中不同尺度的关键特征，弥补了CNN在多尺度特征提取上的不足。为验证M-UNet的性能，本文在Synapse数据集和ACDC数据集上进行了实验。实验结果表明，相较于传统基于CNN的分割方法，M-UNet在分割精度和计算速度两方面均有显著提升。

Abstract: In the architecture of medical image segmentation networks, Convolutional Neural Networks (CNNs) are undoubtedly the most powerful networks, achieving significant breakthroughs due to their strong local feature extraction capabilities, making them one of the mainstream methods in this field. However, the inherent local perception mechanism of CNNs imposes clear limitations when capturing long-distance dependency information in images, which somewhat restricts their segmentation performance. Conversely, Transformer models, with their self-attention mechanism, exhibit significant advantages in handling long-distance dependencies, enabling effective exploration of the feature relationships across the entire image. However, the computational complexity of the self-attention mechanism increases quadratically with the length of the input sequence, leading to issues of excessively high computational costs and large memory usage when dealing with high-resolution medical images. To address this issue, this paper proposes a new segmentation network—M-UNet. The innovation of this network lies in the combination of the unique VSS block structure of the Mamba model with the dilated spatial convolution pyramid pooling module. By leveraging the linear complexity characteristic of the VSS block, it efficiently models global information while significantly reducing computational memory. Additionally, utilizing the multi-scale dilated convolution advantages of the ASPP module allows for precise capture of key features at different scales within medical images, compensating for CNNs’ shortcomings in multi-scale feature extraction. To validate the performance of M-UNet, experiments were conducted on the Synapse dataset and the ACDC dataset. The experimental results indicate that compared to traditional CNN-based segmentation methods, M-UNet shows significant improvements in both segmentation accuracy and computational speed.