Fast motion-compensated reconstruction for 4D-CBCT using deep learning-based groupwise registration.

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Deep learning (DL) enhanced 4D cone beam computed tomography (4D-CBCT) improves motion modeling and motion-compensated (MoCo) reconstruction.
• Conventional deformable image registration (DIR) for motion modeling at treatment time is not temporally feasible.
• There is a need for efficient DL-based registration methods to rapidly generate motion models for 4D-CBCT prior to treatment.

Purpose of the Study:

• To enhance the efficiency of 4D-CBCT MoCo reconstruction using DL-based registration.
• To rapidly generate a motion model prior to treatment using DL-based methods.
• To evaluate the accuracy and efficiency of DL-based DIR models compared to conventional methods.

Main Methods:

• An artifact-reduction DL model was applied to improve initial 4D-CBCT reconstructions.
• Groupwise DL-based DIR was employed to estimate inter-phase motion models.
• Two DL DIR models (patient-specific and population-based) were compared against conventional Elastix DIR using multiple datasets.

Main Results:

• DL DIR models achieved registration accuracy comparable to state-of-the-art conventional methods.
• Final MoCo reconstruction image quality was not significantly different between DL and conventional approaches.
• Average MoCo reconstruction runtime was drastically reduced: from 01:37:26 (conventional) to 00:10:59 (patient-specific DL) and 00:01:05 (population DL).

Conclusions:

• DL-based registration methods significantly improve the efficiency of generating motion models for 4D-CBCT.
• These DL methods do not compromise the performance or image quality of the final MoCo reconstruction.
• DL-based registration offers a viable solution for rapid, accurate motion modeling in 4D-CBCT radiotherapy.