文献类型：期刊文献

英文题名：A Recursive Deformable Mamba Pyramid Network With Multi-Objective Constraints and Spatial-Aware Attention Enhancement for Unimodal and Multimodal Brain Image Registration

作者：Shi, Yupeng[1];Wang, Zichen[1];Wang, Yu[1];Bao, Caiyin[1];Yang, Jianlan[2]

第一作者：史玉攀

通信作者：Yang, JL[1]

机构：[1]Gansu Univ Tradit Chinese Med, Sch Med Informat Engn, Lanzhou 730000, Gansu, Peoples R China;[2]Quanzhou Orthoped Traumatol Hosp, Quanzhou 362000, Fujian, Peoples R China

第一机构：甘肃中医药大学

通信机构：[1]corresponding author), Quanzhou Orthoped Traumatol Hosp, Quanzhou 362000, Fujian, Peoples R China.

年份：2026

卷号：14

起止页码：97803

外文期刊名：IEEE ACCESS

收录：;EI(收录号：20262721026016);WOS:【SCI-EXPANDED(收录号:WOS:001811613700004)】；

基金：This work was supported in part by Quanzhou Science and Technology Plan Project of Fujian Province under Grant 2026QZNY048.

语种：英文

外文关键词：Deformation; Modeling; Modules (abstract algebra); Image registration; Biomedical imaging; Convolution; Accuracy; Computers; Strontium; Labeling; Deformable image registration; mamba; recursive pyramid; spatial semantic awareness; multi-object constraints; brain MRI-CT

摘要：Deformable registration is a critical technology in medical image analysis, playing an indispensable role in various unimodal and multimodal image alignment tasks. Current brain image registration faces two major challenges: firstly, the modeling and fine-grained registration capabilities for complex anatomical deformations of the brain are limited; secondly, the high resource overhead of mainstream Transformer architectures due to self-attention computation makes efficient deployment in standard clinical equipment difficult. To address these issues, this paper proposes a lightweight recursive deformable Mamba pyramid registration network for both unimodal and multimodal brain registration tasks. Specifically, we constructed a novel recursive pyramid framework that fully leverages multi-granularity hierarchical feature collaborative processing to effectively improve registration accuracy. On one hand, we innovatively introduce a dilated convolutional Mamba module at the encoder. While maintaining low resource overhead, this not only effectively alleviates error accumulation between pyramid levels but also achieves efficient processing of high-resolution images and accurate capture of long-range spatial dependencies. On the other hand, at the decoder, we design a multi-head local attention module that integrates spatial semantic awareness to generate the deformation field by fusing deep semantic information. Furthermore, we designed a composite loss function with multi-objective constraints. Its anti-folding penalty term effectively suppresses topological folding in the deformation field, ensuring the differential homeomorphism and physiological plausibility of the registration results. We conducted comprehensive evaluations on three public datasets (LPBA40, IXI, and SR-Reg). Experimental results demonstrate that the proposed method outperforms state-of-the-art methods across key metrics. Specifically, it yields a maximum Dice similarity coefficient improvement of 5.3% for inter-patient registration, achieves a Dice of 75.1% for atlas-to-patient registration, and a near-zero folding voxel rate(0.0004%), while also exhibiting superior multimodal alignment accuracy. Even for data without affine pre-alignment, the proposed network exhibits robust performance when handling large deformations.