基于嵌套U型的3D脑MRI配准网络  

3D Brain MRI Registration Network Based on Nested U-Shapes

作　　者：孙克雷[1] 童波 潘宇 SUN Ke-lei;TONG Bo;PAN Yu(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232000,Anhui,China;School of Artificial and Intelligence,Anhui University of Science and Technology,Huainan 232000,Anhui,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232000 [2]安徽理工大学人工智能学院,安徽淮南232000

出　　处：《兰州文理学院学报(自然科学版)》2025年第1期53-58,共6页Journal of Lanzhou University of Arts and Science(Natural Sciences)

基　　金：安徽省重点研究与开发计划(202004b11020029);国家自然科学基金项目(61703005)。

摘　　要：针对传统的基于U-Net的图像配准网络对图像细节信息提取不够精确的问题,引入了一种嵌套的U型配准网络,在网络的每级编码解码阶段引入RSU-L的U型残差模块,并且在U型网络中的解码阶段每上采样一次均生成配准场,最后将配准场进行叠加得到最终的配准形变场.在公开数据集中,与传统的ANTs、Voxelmorph和最新的Transmorph网络相比,提出的嵌套U型网络在Dice系数上提升了1%~11%,增加了网络模型在图像配准任务上的精确度,对于临床诊断具有一定的帮助.To address the issue of inaccurate extraction of image details in traditional U-Net based image registration networks,a nested U-shaped registration network is introduced.The U-shaped residual module of RSU-L is introduced in each encoding and decoding stage of the network,and a registration field is generated every time the decoding stage in the U-shaped network is upsampled.Finally,the registration field is superimposed to obtain the final registration deformation field.In open datasets,compared with traditional ANTs,Voxelmorph,and the latest Transmorph networks,the proposed nested U-shaped network improves the Dice coefficient by 1% to 11%,increased the accuracy of network models for image registration tasks and providing some assistance for clinical diagnosis and treatment.

关 键 词：图像配准 嵌套结构 U型网络 配准场融合 残差结构 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]