文献类型：期刊文献

中文题名：基于多层级特征融合的多模态医学图像配准

英文题名：Multimodal medical image registration based on multi-layer feature fusion

作者：常青[1];李梦珂[1];陆晨豪[1];张扬[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237

年份：2025

卷号：42

期号：5

起止页码：645

中文期刊名：中国科学院大学学报(中英文)

外文期刊名：Journal of University of Chinese Academy of Sciences

收录：;北大核心:【北大核心2023】；

基金：国家自然科学基金(61976091)资助。

语种：中文

中文关键词：多层级特征融合;多模态;密集对称尺度不变特征变换;无监督配准

外文关键词：multi-layer feature fusion;multimodal;dense symmetric scale invariant feature transform;unsupervised registration

摘要：多模态医学图像的灰度和纹理结构差别较大,难以提取相对应的特征,导致配准精度较低。针对这一问题,提出基于多层级特征融合的配准模型,并行提取参考图和浮动图的特征,在多层级结构中使用双输入空间注意力模块实现多模态特征渐进融合,获取其相关性,并将这种相关性映射到图像配准变换中。同时使用基于密集对称尺度不变特征变换的局部特征相似性引导网络进行迭代优化,实现多模态图像的无监督配准。
As the initial step of multimodal medical image registration,the accuracy and speed of registration will largely affect the effect of medical image fusion.Due to the large difference in grayscale and texture structure of multimodal medical images,it is difficult to extract correlating features,resulting in low registration accuracy.This paper proposes a multi-layer feature fusion registration network,parallel extraction of features of the fix image and moving image,and the multimodal feature is gradually fused by using the dual-input spatial attention module in the multilayer structure,obtaining their correlation and mapping such correlation to image registration transformation.At the same time,the structural information loss term guidance network based on dense symmetric scale invariant feature transform is introduced for iterative optimization to achieve accurate unsupervised registration.