摘要
针对三维图像配准存在易丢失空间信息、拓扑结构保持困难及耗时长等问题,提出一种混合的Transformer-ConvNet模型。在经典VoxelMorph模型基础上,引入了交叉注意力模块,进行有效的远程建模和高维数据处理;针对运算量大的问题,在卷积层引入了inception模块,采用并行连接的方式将多个不同尺寸和不同类型的卷积层相互融合,学习细粒度特征,以提高配准精度。在脑部MR数据集上,进行了消融实验,结果表明,引入交叉注意力机制和inception模块网络较VoxelMorph,Dice系数提高了11.5%。将所提算法与4种经典算法进行了对比,结果表明,所提模型配准性能有了明显提升,较VoxelMorph、CycleMorph、ViT-V-Net、TransMorph分别提高9.6%、8.7%、9.1%、6.3%,且参数量较少。
For 3D image registration,there are some problems such as easy loss of spatial information,being difficult to effectively maintain topological structure and time-consuming.Therefore,a mixed Transformer-ConvNet model is proposed.Based on the classical VoxelMorph model,cross-attention module is introduced to carry out effective remote modeling and high-dimensional data processing;to solve the problem of large amount of computation,inception module is introduced into the convolutional layer,which uses parallel connection to integrate multiple convolution layers of different sizes and types to learn fine-grained features to improve registration accuracy.Ablation experiments are performed on the brain MR Dataset,and the results show that the Dice score is improved by 11.5%compared with the VoxelMorph model by introducing the cross-attention mechanism and inception module network.The proposed algorithms are compared with four classical algorithms.The result show that the registration performance of the proposed model is significantly improved,which is 9.6%,8.7%,9.1%and 6.3%higher than that of theVoxelMorph,the CycleMorph,the ViT-V-Net and the TransMorph,respectively,and the number of parameters is small.
作者
杨丽丽
王玉
王明泉
商奥雪
崔瑞杰
商然
YANG Lili;WANG Yu;WANG Mingquan;SHANG Aoxue;CUI Ruijie;SHANG Ran(School of Information and Communication Engineering,North University of China,Taiyuan 030051,China)
出处
《机械与电子》
2024年第11期11-16,共6页
Machinery & Electronics
基金
山西省重点研发计划资助项目(201803D121069)
山西省应用基础研究项目面上自然基金项目(201801D121162)。