郑州大学学报（工学版）2024，Vol.45Issue(2)：27-32,6.DOI:10.13705/j.issn.1671-6833.2023.05.007
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
摘要
Abstract
For medical image segmentation tasks such as glioma image segmentation with dense prediction,both lo-cal 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 ser-ial 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 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Key words
deep learning/brain glioma image segmentation/CNN/Transformer/self-attention mechanism分类
数理科学引用本文复制引用
高宇飞,马自行,徐静,赵国桦,石磊..基于卷积和可变形注意力的脑胶质瘤图像分割[J].郑州大学学报（工学版）,2024,45(2):27-32,6.基金项目
国家自然科学基金资助项目(62006210) （62006210）
河南省重大公益专项(201300210500) （201300210500）
郑州大学高层次人才科研启动基金(32340306) （32340306）
