An unsupervised 2D-3D deformable registration network (2D3D-RegNet) for cone-beam CT estimation.

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Limited-angle cone-beam CT (CBCT) acquisition offers reduced imaging time and dose, crucial for real-time target localization during arc-based radiotherapy.
• However, undersampled angular data in limited-angle CBCT leads to significant image distortions and artifacts, hindering clinical utility.
• Traditional 2D-3D deformable registration methods, while effective, are computationally intensive and time-consuming, posing a bottleneck for rapid clinical application.

Purpose of the Study:

• To develop a fast and accurate unsupervised, end-to-end deep learning framework for 2D-3D deformable registration to address the computational limitations of conventional methods.
• To evaluate the performance of the developed framework in generating high-quality deformation vector fields (DVFs) and CBCT images from limited-angle projections.
• To assess the robustness and potential clinical integration of the framework, including its synergy with biomechanical modeling.

Main Methods:

• Development of a novel convolutional neural network (CNN) based framework, termed 2D3D-RegNet, for unsupervised 2D-3D deformable image registration.
• Utilizing limited-angle cone-beam projections to estimate the deformation vector field (DVF) and reconstruct a new CBCT volume.
• Assessing DVF accuracy against 3D-3D registration and conventional 2D-3D registration, and performing robustness analysis against variations in projection sampling and angular offsets.

Main Results:

• The 2D3D-RegNet framework successfully generated accurate DVFs within 5 seconds for 90 projections covering a 90° scan angle.
• Achieved DVF accuracy superior to 3D-3D deformable registration and comparable to conventional iterative 2D-3D methods.
• Demonstrated the framework's ability to suppress artifacts and distortions in reconstructed CBCT images, reflecting updated patient anatomy.
• Preliminary analysis indicated robustness to variations in angular sampling frequency and scan angle offsets.

Conclusions:

• The developed 2D3D-RegNet offers a computationally efficient solution for 2D-3D deformable registration in limited-angle CBCT, overcoming the speed limitations of traditional iterative algorithms.
• The framework shows promise for improving the quality and clinical applicability of CBCT in radiotherapy by enabling rapid, artifact-free image reconstruction.
• 2D3D-RegNet can serve as a fast DVF computation core, potentially integrated with biomechanical models for further refinement and enhanced clinical workflow.