Related Experiment Video
Updated: Jun 17, 2026
05:05
Four-Dimensional CT Analysis Using Sequential 3D-3D Registration
Published on: November 23, 2019
8.4K
A hybrid population-based and patient-specific framework for 2D-3D deformable registration-driven limited-angle
Xiaoxue Qian1, Hua-Chieh Shao1, You Zhang1
1The Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, Texas, USA.
Medical Physics
|October 28, 2025
Summary
A novel hybrid framework (HB-2D3DReg) improves limited-angle cone-beam CT (LA-CBCT) estimation by combining population-based and patient-specific 2D-3D deformable registration. This approach enhances accuracy and efficiency for anatomy monitoring in radiotherapy.
Area of Science:
- Medical Imaging
- Computational Anatomy
- Radiotherapy Physics
Background:
- Limited-angle cone-beam CT (LA-CBCT) offers reduced imaging time and dose but is prone to severe artifacts due to under-sampling.
- 2D-3D deformable registration is a key technique for mitigating LA-CBCT artifacts by deforming prior CT/CBCT data.
- Existing methods like population-trained networks offer fast inference but limited accuracy, while patient-specific models are adaptable but slow.
Purpose of the Study:
- To develop a hybrid 2D-3D deformable registration framework (HB-2D3DReg) that enhances both accuracy and efficiency in LA-CBCT estimation.
- To synergize the strengths of population-based and patient-specific registration models while minimizing their respective drawbacks.
Main Methods:
- A two-stage hybrid framework integrating a population-trained 2D-3D registration network (2D3D-RegNet) and a patient-specific implicit neural representation network (2D3D-INR).
- Unsupervised training of 2D3D-RegNet using a similarity loss between digitally reconstructed radiographs (DRRs) and limited-angle projections.
- Test-time refinement of deformation-vector-fields (DVFs) by the 2D3D-INR network, leveraging the population model for accelerated optimization and improved accuracy.
Main Results:
- HB-2D3DReg demonstrated superior LA-CBCT estimation and registration accuracy compared to existing methods and no registration.
- Achieved a mean image relative error of 7.99 ± 2.16% and target registration error of 3.70 ± 1.94 mm under challenging scan conditions.
- Reduced test-time optimization to approximately 3 minutes, significantly faster than the 13 minutes required for the 2D3D-INR-only approach.
Conclusions:
- The HB-2D3DReg framework provides accurate and robust 2D-3D deformable registration for LA-CBCT estimation.
- Enables efficient and reliable anatomy monitoring crucial for guiding radiotherapy treatments.
- The developed code will be publicly released to facilitate further research and application.
