摘要
肝脏及肝肿瘤在三维计算机断层扫描(CT)影像中的精准分割,对临床诊断与治疗规划至关重要。针对现有卷积神经网络(CNN)分割方法因感受野受限难捕长距离依赖,而Transformer分割方法在样本有限的三维医学影像中应用受限的问题,文章提出基于Mamba的三维医学图像分割新模型。该模型含三维门控空间卷积模块(3DGSCM)与通道混洗上采样模块,分别提取空间关联特征与共性特征。两大模块协同增强模型对全局结构的建模能力,同时保障了精确的边界定位效果。在LiTS2017数据集上的实验结果显示,肝脏分割的Dice相似系数(DSC)达96.42%,肝肿瘤分割DSC达70.70%;在3D-IRCADb数据集上的泛化性实验中,肝脏与肝肿瘤分割DSC分别达到96.79%和67.10%。多项对比实验结果进一步验证了所提模型在分割性能上的优越性与鲁棒性。
Accurate segmentation of liver and liver tumors in three-dimensional Computed Tomography(CT)images is crucial for clinical diagnosis and treatment planning.To address the issues that existing Convolutional Neural Network(CNN)segmentation methods struggle to capture long-range dependencies due to limited receptive fields,and Transformer segmentation methods are restricted in application on three-dimensional medical images with limited samples,this paper proposes a new 3D medical image segmentation model based on Mamba.The model contains a 3D gated spatial convolution module and a channel shuffle upsampling module,which extract spatial correlation features and common features,respectively.The two modules synergistically enhance the modeling ability of the model for global structures,while simultaneously ensuring precise boundary positioning.Experimental results on the LiTS2017 dataset show that the Dice Similarity Coefficient(DSC)of liver segmentation reaches 96.42%,and the DSC of liver tumor segmentation reaches 70.70%.In the generalization experiment on the 3D-IRCADb dataset,the DSCs of liver and liver tumor segmentation reach 96.79%and 67.10%,respectively.Results of multiple comparative experiments further verify the superiority and robustness of the proposed model in segmentation performance.
作者
郭佳豪
胡怀飞
GUO Jiahao;HU Huaifei(South-Central Minzu University,Wuhan 430074,China)
出处
《现代信息科技》
2026年第3期82-87,共6页
Modern Information Technology
