基于卷积和可变形注意力的脑胶质瘤图像分割  

Brain Glioma Image Segmentation Based on Convolution and Deformable Attention

作　　者：高宇飞[1,2] 马自行 徐静 赵国桦[4] 石磊 GAO Yufei;MA Zixing;XU Jing;ZHAO Guohua;SHI Lei(School of Cyber Science and Engineering,Zhengzhou University,Zhengzhou 450002,China;Songshan Laboratory,Zhengzhou 450052,China;School of Computer and Artificial Intelligence,Zhengzhou University,Zhengzhou 450001,China;Department of Magnetic Resonance,The First Affiliated Hospital of Zhengzhou University,Zhengzhou 450003,China)

机构地区：[1]郑州大学网络空间安全学院,河南郑州450002 [2]嵩山实验室,河南郑州450052 [3]郑州大学计算机与人工智能学院,河南郑州450001 [4]郑州大学第一附属医院,河南郑州450003

出　　处：《郑州大学学报（工学版）》2024年第2期27-32,共6页Journal of Zhengzhou University（Engineering Science）

基　　金：国家自然科学基金资助项目(62006210);河南省重大公益专项(201300210500);郑州大学高层次人才科研启动基金(32340306)。

摘　　要：对于脑胶质瘤图像分割这类密集预测的医学影像分割任务,局部和全局依赖关系都是不可或缺的,针对卷积神经网络缺乏建立全局依赖关系的能力,且自注意力机制在局部细节上捕捉能力不足等问题,提出了基于卷积和可变形注意力的脑胶质瘤图像分割方法。设计了卷积和可变形注意力Transformer的串行组合模块,其中卷积用于提取局部特征,紧随其后的可变形注意力Transformer用于捕捉全局依赖关系,建立不同分辨率下局部和全局依赖关系。作为一种CNN-Transformer混合架构,所提方法不需要任何预训练即可实现精准的脑胶质瘤图像分割。实验结果表明:所提方法在BraTS2020脑胶质图像分割数据集上平均Dice系数和平均95%豪斯多夫距离分别为83.56%和11.30 mm,达到了与其他脑胶质瘤图像分割方法相当的分割精度,同时降低了至少50%的计算开销,有效提升了脑胶质瘤图像分割的效率。Accurate and efficient brain glioma image segmentation is important for clinical diagnosis,treatment,and postoperative observation.For medical image segmentation tasks such as glioma image segmentation with dense prediction,both local and global dependencies were indispensable.To address the problems that convolutional neural networks lacked the ability to establish global dependencies and the self-attention mechanism had insufficient ability to capture local details,a convolutional and deformable attention-based method for glioma image segmentation was proposed.A serial combination module of convolution and deformable attention Transformer was designed,in which convolution was used to extract local features and the immediately following deformable attention.Transformer was used to capture global dependencies to the establishment of local and global dependencies at different resolutions.As a hybrid CNN-Transformer architecture,the method could achieve accurate brain glioma image segmentation without any pre-training.Experiments showed that the average dice score and the average 95%Hausdorff distance on the BraTS2020 glioma image segmentation dataset were 83.56%and 11.30 mm,respectively,achieving comparable segmentation accuracy compared with international advanced glioma image segmentation methods,respectively,achieving comparable segmentation accuracy compared with other methods,while reducing the computational overhead by at least 50%and effectively improving the efficiency of glioma image segmentation.

关 键 词：深度学习 脑胶质瘤图像分割 卷积神经网络 TRANSFORMER 自注意力机制 

分 类 号：O244[理学—计算数学] TP391.41[理学—数学]