摘要
多模态MRI脑肿瘤影像的精准分割对脑癌临床诊疗及预后评估至关重要。针对卷积神经网络在捕获全局上下文信息和建立长远程依赖关系方面存在的局限性,提出了基于Mamba与U-Net融合架构的PhC-ToMamba分割模型。模型在瓶颈层嵌入了ToM模块旨在有效建模高维特征的全局信息,通过从三个方向计算特征依赖关系并交互,提取更适用于三维图像的全局特征信息;此外,为进一步提升全局特征的提取能力,提出了一种新的多面体卷积(PhConv),并将其嵌入至编码器中,显著扩大了感受野,并提升对重点目标区域的特征提取能力,有效解决了当前主流脑肿瘤图像分割模型对全局信息感知的局限性问题,增强了对关键区域的关注度。在BraTS 2021和MSD Task01_BrainTumor数据集上进行了广泛的实验。实验结果显示,PhC-ToMamba在整个肿瘤、肿瘤核心和增强肿瘤分割任务中的Dice系数分别达到了95.05%/90.46%、94.53%/89.91%和90.74%/75.91%。与其他先进方法相比,PhC-ToMamba在分割精度和参数效率方面展现了优越性,为脑肿瘤分割任务提供稳健的解决方案,从而提高了诊断准确性。
Accurate segmentation of multimodal MRI brain tumor images is crucial for clinical diagnosis and prognosis assessment of brain cancer.To address the limitations of convolutional neural networks in capturing global contextual information and modeling long-range dependencies,this paper proposed a novel segmentation model named Polyhedron Conv-Tri-orientated Mamba(PhC-ToMamba)by integrating Mamba with a U-Net architecture.It embedded a Tri-orientated Mamba(ToM)modu-le in the bottleneck layer to model high-dimensional global features by computing and interacting dependencies along three directions,thereby enhancing global feature representation in 3D medical images.In addition,this paper introduced a novel Polyhedron Convolution(PhConv)into the encoder to enlarge the receptive field and improved the extraction of critical target regions.These modules effectively enhanced global context awareness and focused attention on key tumor regions.Extensive experiments were conducted on the BraTS 2021 and MSD Task01_BrainTumor datasets.The proposed PhC-ToMamba achieves Dice scores of 95.05%/90.46%,94.53%/89.91%,and 90.74%/75.91%for whole tumor,tumor core,and enhancing tumor segmentation,respectively.Compared with state-of-the-art methods,PhC-ToMamba demonstrates superior segmentation accuracy and parameter efficiency,providing a robust solution for brain tumor segmentation and improving diagnostic precision.
作者
张野
牛大田
Zhang Ye;Niu Datian(School of Science,Dalian Minzu University,Dalian Liaoning 116000,China)
出处
《计算机应用研究》
北大核心
2026年第1期305-312,共8页
Application Research of Computers
基金
国家自然科学基金资助项目(11872145)
辽宁省教育厅基本科研资助项目(JYTMS20231805)。
