机械与电子2025，Vol.43Issue(10)：11-17,7.
脑部MRI图像配准中CNN与Transformer并行架构的算法研究
Research on a CNN-Transformer Parallel Architecture for Brain MRI Image Registration
摘要
Abstract
To address the current limitations of deep learning-based image registration models,such as the difficulty in effectively leveraging the complementary strengths of CNN and Transformers,limited registration accuracy,and challenges in preserving the topological structure of original images,we propose an unsupervised CNN-Transformer hybrid registration network.The model is built using the Swin-Transformer,known for its state-of-the-art registration accuracy,and a lightweight CNN architecture.By fusing the extracted features,the model combines the local feature extraction capabilities of CNN with the global modeling strengths of Transformers,resulting in more accurate and lightweight registration.We evaluated the network on two public brain MRI datasets(IXI and LPBA40).Experimental results demon-strate that our model significantly outperforms VoxelMorph,Pvt,ViT-V-Net,and TransMorph in met-rics such as DICE and structural similarity,while maintaining the efficiency advantages of learning-based methods,showcasing superior registration performance.关键词
深度学习/Transformer架构/图像配准/卷积神经网络Key words
deep learning/Transformer architecture/image registration/convolutional neural network分类
计算机与自动化引用本文复制引用
刘源,王玉,杨丽丽,杨洁,张宇昊..脑部MRI图像配准中CNN与Transformer并行架构的算法研究[J].机械与电子,2025,43(10):11-17,7.基金项目
山西省应用基础研究项目面上自然基金项目(201801D121162) （201801D121162）
山西省重点研发计划资助项目(201803D121069) （201803D121069）
中北大学重点实验室开发研究基金资助项目(DXMBJJ2024-04) （DXMBJJ2024-04）
