Author: Yunfei Dong, Dongyang Guo, Zhenyu Yang, Fang-Fang Yin, Zeyu Zhang 👨🔬
Affiliation: Duke University, Duke Kunshan University, Medical Physics Graduate Program, Duke Kunshan University 🌍
Purpose:
To develop a Biomechanically Guided Deep Learning Registration Network (BG-DRNet) that improves both accuracy and physiological plausibility in liver image registration. While cone-beam CT (CBCT) offers real-time 3D imaging in the treatment room, it suffers from limited soft-tissue contrast and artifacts. MRI provides superior soft-tissue contrast but is less readily available due to higher costs. By integrating MRI and CBCT via advanced registration, we aim to enhance tumor localization for improved treatment outcomes. Additionally, the framework is designed to address single-modality (MR-to-MR) registration with the same network architecture.
Methods:
We introduce a hybrid liver image registration approach that combines finite element modeling with deep learning to handle both multi-modality (MR-CBCT) and single-modality (MR-MR) registration tasks. Using Finite Element Method (FEM) simulations, we capture respiratory-induced liver deformations and generate deformation vector fields (DVFs) to guide registration. Our BG-DRNet employs a 3D self-attention encoder–decoder architecture to learn complex inter-modality relationships and predict physiologically plausible deformations. This approach leverages the physical realism of FEM with the computational efficiency of deep learning in an end-to-end solution capable of multiple registration scenarios.
Results:
We used landmark localization of liver vessels as the primary evaluation criterion to assess anatomical fidelity. The proposed biomechanically guided deep learning approach demonstrated superior registration accuracy and physiological plausibility in multi-modality (MR-CBCT) scenarios compared to traditional deep learning-based methods. In addition, it maintained robust performance for single-modality (MR-MR) registration without altering the underlying model structure. Notably, it also significantly reduced the patient-specific computation time typically required for standalone finite element analyses.
Conclusion:
Our findings indicate that the biomechanically guided deep learning registration network substantially enhances both registration accuracy and physiological relevance for multi-modality and single-modality tasks. This improvement has the potential to advance clinical liver image registration applications, paving the way for more precise and efficient treatment strategies.