TY - JOUR
T1 - GraphMorph
T2 - Equilibrium adjustment regularized dual-stream GCN for 4D-CT lung imaging with sliding motion
AU - Yuan, Peng
AU - Lyu, Fei
AU - Zhang, Yudong
AU - Yang, Chunfeng
AU - Gao, Zhiqiang
AU - Wu, Zhan
AU - Dong, Jianmin
AU - Lyu, Tianling
AU - Zhao, Wei
AU - Coatrieux, Jean Louis
AU - Chen, Yang
N1 - This work was supported in part by the State Key Project of Research and Development Plan under Grants 2022YFC2408500, the Science and Technology Project of Xizang Autonomous Region under Grant XZ202401JD0009, the National Natural Science Foundation of China under Grant T2225025, 62401141 and 62401512, and the Key Research and Development Programs in Jiangsu Province of China under Grants BE2021703 and BE2022768, and National Natural Science Foundation of China under Grant 12175012, and the Jiangsu Province Science Foundation for Youths under Grant BK20241305, and China Postdoctoral Science Foundation under Grant 2023M740607. This research work is supported by the Big Data Computing Center of Southeast University.
Publisher Copyright:
© 2025 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
PY - 2025/11/12
Y1 - 2025/11/12
N2 - Four-dimensional computed tomography (4D-CT) is crucial for radiation therapy, enabling motion tracking of thoracic and abdominal tumors. However, respiratory-induced sliding motion at organ interfaces poses a significant challenge for existing registration algorithms, leading to difficulties in accurately modeling discontinuous deformations. We propose an Equilibrium Adjustment Regularized Dual-Stream Graph Convolutional Network (GCN), dubbed GraphMorph, for addressing the challenges of sliding motion in 4D-CT lung registration. This network ensures physically consistent deformations while balancing smoothness and discontinuity at sliding interfaces. To address the challenge of global large deformations, we propose a Topology Enhanced Graph Attention (TEGA) module, which harnesses the topological information of the graph structure and the global dependencies modeled by the Transformer, facilitating the learning of regional relationships in sliding organs. We also construct a Cross-Scale Contextual Aggregation (CSCA) module that aggregates contextual information by leveraging content correlations between images from different respiratory phases, thus addressing subtle local deformations. To balance continuous and discontinuous deformations, we designed a biomechanics-inspired Equilibrium Adjustment Regularization (EAR) method, which eliminates discretization dependencies and effectively mitigates the impact of local intensity inhomogeneities on registration accuracy. Experiments on public and in-house 4D datasets demonstrate that GraphMorph achieves an average target registration error (TRE) of 0.96 mm, outperforming existing methods even on out-of-distribution (OOD) data. GraphMorph enhances 4D-CT lung imaging by accurately modeling sliding motion, enabling precise deformable registration for lesion localization in image-guided interventions. The source code is available at https://github.com/computerAItest/GraphMorph.
AB - Four-dimensional computed tomography (4D-CT) is crucial for radiation therapy, enabling motion tracking of thoracic and abdominal tumors. However, respiratory-induced sliding motion at organ interfaces poses a significant challenge for existing registration algorithms, leading to difficulties in accurately modeling discontinuous deformations. We propose an Equilibrium Adjustment Regularized Dual-Stream Graph Convolutional Network (GCN), dubbed GraphMorph, for addressing the challenges of sliding motion in 4D-CT lung registration. This network ensures physically consistent deformations while balancing smoothness and discontinuity at sliding interfaces. To address the challenge of global large deformations, we propose a Topology Enhanced Graph Attention (TEGA) module, which harnesses the topological information of the graph structure and the global dependencies modeled by the Transformer, facilitating the learning of regional relationships in sliding organs. We also construct a Cross-Scale Contextual Aggregation (CSCA) module that aggregates contextual information by leveraging content correlations between images from different respiratory phases, thus addressing subtle local deformations. To balance continuous and discontinuous deformations, we designed a biomechanics-inspired Equilibrium Adjustment Regularization (EAR) method, which eliminates discretization dependencies and effectively mitigates the impact of local intensity inhomogeneities on registration accuracy. Experiments on public and in-house 4D datasets demonstrate that GraphMorph achieves an average target registration error (TRE) of 0.96 mm, outperforming existing methods even on out-of-distribution (OOD) data. GraphMorph enhances 4D-CT lung imaging by accurately modeling sliding motion, enabling precise deformable registration for lesion localization in image-guided interventions. The source code is available at https://github.com/computerAItest/GraphMorph.
KW - 4D-CT lung registration
KW - Graph convolutional network
KW - Regularization term
KW - Sliding motion
KW - Topological information
UR - https://www.scopus.com/pages/publications/105022475221
U2 - 10.1016/j.neucom.2025.132022
DO - 10.1016/j.neucom.2025.132022
M3 - Journal article
AN - SCOPUS:105022475221
SN - 0925-2312
VL - 664
JO - Neurocomputing
JF - Neurocomputing
M1 - 132022
ER -
