A multi-stage CT-MRI registration method for spine based on multi-vertebra structural modeling

doi:10.11996/JG.j.2095-302X.2026051033

Abstract

Abstract:

In minimally invasive spinal surgery, a single imaging modality cannot simultaneously provide sufficient information on both osseous structures and soft tissues, making intraoperative spatial localization challenging. Conventional cross-modality registration methods often struggle to accommodate the intensity differences between CT and MRI, inconsistent scan coverage, imaging artifacts, and the distinct motion characteristics of multiple spinal tissues, which may lead to registration failure and vertebral localization errors. To address these issues, a multi-stage spinal CT-MRI registration method based on multi-vertebra structural modeling was proposed. In the coarse registration stage, anatomical constraints were constructed using vertebral segmentation results. Common vertebral labels were extracted from the two modalities, and a representative vertebra was selected as an anchor to initialize rigid and affine transformations, thereby reducing large-scale pose differences and preventing non-rigid registration from becoming trapped in local optima. In the fine registration stage, a deep neural network integrating multi-vertebra affine modeling and diffeomorphic deformation was developed. The multi-vertebra affine branch characterized the approximately rigid motion of different vertebral segments, while velocity-field integration was employed to generate a diffeomorphic deformation field for modeling the nonlinear deformation of intervertebral discs and surrounding soft tissues, thereby achieving high-precision non-rigid registration. Experiments were conducted on a self-built spinal CT-MRI dataset. The proposed method achieved Dice, HD₉₅, and ASD values of 91.43%, 2.61 mm, and 1.16 mm, respectively, in the lumbar region, outperforming all comparison methods. Quantitative results demonstrated that the proposed method improved the regional overlap and boundary alignment accuracy of lumbar vertebral structures. The visualization results further demonstrated that it could effectively coordinate the local near-rigid motion of vertebral segments with the continuous deformation of surrounding tissues, maintaining good stability and anatomical rationality. This provided a technical foundation for preoperative multimodal image fusion and subsequent surgical navigation.

Key words: spine, multimodal registration, CT-MRI, deep learning, diffeomorphic registration

CLC Number:

YANG Zhiyong, LIU Di, JIANG Shan, ZHOU Zeyang, HAN Jun, ZHU Shan, CHEN Chao. A multi-stage CT-MRI registration method for spine based on multi-vertebra structural modeling[J]. Journal of Graphics, 2026, 47(5): 1033-1042.

Figures/Tables 10

References 23

[1]	JITPAKDEE K, LIU Y T, HEO D H, et al. Minimally invasive endoscopy in spine surgery: where are we now?[J]. European Spine Journal, 2023, 32(8): 2755-2768. DOI PMID
[2]	MAAYAN O, MAI E, KIM A Y E, et al. Overview of endoscopic spine surgery and learning curve[J]. Seminars in Spine Surgery, 2024, 36(1): 101079. DOI URL
[3]	EL CHOUEIRI J, PELLICANÒ F, CAIMI E, et al. Advancements in spinal endoscopic surgery: comprehensive techniques and pathologies addressed by full endoscopy beyond lumbar disc herniation[J]. Journal of Clinical Medicine, 2025, 14(11): 3685. DOI URL
[4]	GUO X B, JIN J T, CHEN J W, et al. Preoperative application of CT and MRI registration in lumbar disc herniation endoscopic surgery could improve the postoperative rehabilitation of patients[J]. American Journal of Translational Research, 2024, 16(6): 2453-2463. DOI PMID
[5]	LI H Y, LIU H L, XIAO M Q, et al. CT-3D MERGE fusion imaging improves image quality compared with CT and 3D MERGE in patients with lumbar disc herniation[J]. Frontiers in Neurology, 2024, 15: 1490033. DOI URL
[6]	JIANG Z W, ZHU Y H, WANG W, et al. Evaluation of MRI-based synthetic CT for lumbar degenerative disease: a comparison with CT[J]. Scientific Reports, 2025, 15(1): 20548. DOI
[7]	KAVRAKOVA I G, HAAGE P, STUECKLE C A. Current state-of-the-art 3D MRI sequences for assessing bone morphology with emphasis on cranial and spinal imaging: a narrative review[J]. RöFo: Fortschritte auf dem Gebiet der Röntgenstrahlen und der bildgebenden Verfahren, 2026, 198(3): 301-324. DOI URL
[8]	MOK T C W, CHUNG A C S. Fast symmetric diffeomorphic image registration with convolutional neural networks[C]// 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2020: 4643-4652.
[9]	VIOLA P, WELLS W M. Alignment by maximization of mutual information[C]// IEEE International Conference on Computer Vision. New York: IEEE Press, 1995: 16-23.
[10]	AVANTS B B, TUSTISON N J, SONG G, et al. A reproducible evaluation of ANTs similarity metric performance in brain image registration[J]. NeuroImage, 2011, 54(3): 2033-2044. DOI PMID
[11]	VELESACA H O, BASTIDAS G, ROUHANI M, et al. Multimodal image registration techniques: a comprehensive survey[J]. Multimedia Tools and Applications, 2024, 83(23): 63919-63947. DOI
[12]	DE VOS B D, BERENDSEN F F, VIERGEVER M A, et al. A deep learning framework for unsupervised affine and deformable image registration[J]. Medical Image Analysis, 2019, 52: 128-143. DOI PMID
[13]	BALAKRISHNAN G, ZHAO A, SABUNCU M R, et al. VoxelMorph: a learning framework for deformable medical image registration[J]. IEEE Transactions on Medical Imaging, 2019, 38(8): 1788-1800. DOI URL
[14]	KREBS J, DELINGETTE H, MAILHÉ B, et al. Learning a probabilistic model for diffeomorphic registration[J]. IEEE Transactions on Medical Imaging, 2019, 38(9): 2165-2176. DOI PMID
[15]	CHEN J Y, FREY E C, HE Y F, et al. TransMorph: transformer for unsupervised medical image registration[J]. Medical Image Analysis, 2022, 82: 102615. DOI URL
[16]	JIAN B L, AZAMPOUR M F, DE BENETTI F, et al. Weakly-supervised biomechanically-constrained CT/MRI registration of the spine[C]// The 25th International Conference on Medical Image Computing and Computer Assisted Intervention. Cham: Springer, 2022: 227-236.
[17]	KETCHA M D, DE SILVA T, UNERI A, et al. Multi-stage 3D-2D registration for correction of anatomical deformation in image-guided spine surgery[J]. Physics in Medicine & Biology, 2017, 62(11): 4604-4622.
[18]	STRITTMATTER A, SCHAD L R, ZÖLLNER F G. Deep learning-based affine medical image registration for multimodal minimal-invasive image-guided interventions - a comparative study on generalizability[J]. Zeitschrift für Medizinische Physik, 2024, 34(2): 291-317. DOI URL
[19]	JENA R, SETHI D, CHAUDHARI P, et al. Deep learning in medical image registration: magic or mirage?[C]// The 38th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2024: 108331-108353.
[20]	DIAZ-PINTO A, ALLE S, NATH V, et al. MONAI Label: a framework for AI-assisted interactive labeling of 3D medical images[J]. Medical Image Analysis, 2024, 95: 103207. DOI URL
[21]	AKINCI D’ANTONOLI T, BERGER L K, INDRAKANTI A K, et al. TotalSegmentator MRI: robust sequence-independent segmentation of multiple anatomic structures in MRI[J]. Radiology, 2025, 314(2): e241613. DOI URL
[22]	MATTES D, HAYNOR D R, VESSELLE H, et al. Nonrigid multimodality image registration[C]//Medical Imaging 2001: Image Processing. Bellingham: SPIE, 2001, 4322: 406-417.
[23]	MOK T C W, CHUNG A C S. Large deformation diffeomorphic image registration with Laplacian pyramid networks[C]// The 23rd International Conference on Medical Image Computing and Computer Assisted Intervention. Cham: Springer, 2020: 211-221.

模块	尺度变化	主要操作	通道数
输入层	1	CT与MRI通道拼接	2
编码器	1→1/16	4个3D下采样卷积块，步长为2，InstanceNorm3D，LeakyReLU(0.2)	2→16→32→32→32
多椎体仿射分支	1/16	深层特征展平后分别输入K个全连接层，每层输出一个3×4仿射矩阵	K×12
多尺度仿射先验	1/16，1/8，1/4，1/2	将K个椎体仿射矩阵转换为多尺度稠密位移先验，并在通道维度拼接	3K
解码融合模块	1/16→1	4个组合单元逐层融合编码器特征、上采样特征和仿射先验	32
特征细化模块	1	3个3D卷积块进一步细化全分辨率特征	32→32→16→16
速度场预测层	1	3×3×3卷积输出三维速度场	16→3
形变场生成	1	scaling-and-squaring积分，积分步数为7	3

模块	尺度变化	主要操作	通道数
输入层	1	CT与MRI通道拼接	2
编码器	1→1/16	4个3D下采样卷积块，步长为2，InstanceNorm3D，LeakyReLU(0.2)	2→16→32→32→32
多椎体仿射分支	1/16	深层特征展平后分别输入K个全连接层，每层输出一个3×4仿射矩阵	K×12
多尺度仿射先验	1/16，1/8，1/4，1/2	将K个椎体仿射矩阵转换为多尺度稠密位移先验，并在通道维度拼接	3K
解码融合模块	1/16→1	4个组合单元逐层融合编码器特征、上采样特征和仿射先验	32
特征细化模块	1	3个3D卷积块进一步细化全分辨率特征	32→32→16→16
速度场预测层	1	3×3×3卷积输出三维速度场	16→3
形变场生成	1	scaling-and-squaring积分，积分步数为7	3

方法	Dice /%↑	HD₉₅ /mm↓	ASD /mm↓
Affine^[22]	79.64±6.95	7.58±2.31	3.64±0.81
ANTs-SyN^[10]	81.37±3.34	6.18±1.54	2.37±0.43
VoxelMorph^[13]	88.95±2.05	5.47±1.22	1.61±0.35
LapIRN^[23]	86.24±3.52	5.27±1.47	1.75±0.58
TransMorph^[15]	90.65±1.84	3.85±1.18	1.43±0.38
Ours	91.43±1.52	2.61±0.81	1.16±0.41

方法	Dice /%↑	HD₉₅ /mm↓	ASD /mm↓
Affine^[22]	79.64±6.95	7.58±2.31	3.64±0.81
ANTs-SyN^[10]	81.37±3.34	6.18±1.54	2.37±0.43
VoxelMorph^[13]	88.95±2.05	5.47±1.22	1.61±0.35
LapIRN^[23]	86.24±3.52	5.27±1.47	1.75±0.58
TransMorph^[15]	90.65±1.84	3.85±1.18	1.43±0.38
Ours	91.43±1.52	2.61±0.81	1.16±0.41

方法	配准时间/s	参数量/M	GPU memory/MB
VoxelMorph^[13]	0.496±0.003	0.31	1 682
LapIRN^[23]	0.429±0.003	1.26	2 742
TransMorph^[15]	0.645±0.004	46.73	6 704
Ours	0.710±0.005	8.28	4 733