Deep learning-based motion compensation for four-dimensional cone-beam computed tomography (4D-CBCT) reconstruction.

Area of Science:

• Medical Imaging
• Image Reconstruction
• Deep Learning

Background:

• Motion-compensated (MoCo) reconstruction enhances 4D-CBCT image quality.
• Accurate motion information is crucial for MoCo, but initial 4D-CBCT image quality limits motion modeling.
• Existing methods struggle with artifacts in initial images, hindering data-driven MoCo approaches.

Purpose of the Study:

• Develop a deep learning (DL) method for high-quality motion models in MoCo reconstruction.
• Improve the final image quality of 4D-CBCT through enhanced motion modeling.
• Address limitations of current MoCo techniques by improving initial image quality.

Main Methods:

• Proposed a 3D artifact-reduction convolutional neural network (CNN) to refine conventional PCF reconstructions.
• The CNN mitigates undersampling artifacts while preserving motion information.
• CNN-enhanced images were used for motion modeling in MoCo reconstruction (CNN+MoCo).

Main Results:

• The CNN effectively reduced streaking artifacts in PCF CBCT images across all tested datasets.
• CNN+MoCo reconstruction yielded improved image detail compared to PCF and conventional MoCo.
• Experiments showed significantly improved motion model accuracy using CNN-enhanced images.
• CNN+MoCo demonstrated lower RMSE and higher NCC than other methods on XCAT and SPARE datasets.

Conclusions:

• CNN-based artifact reduction substantially improves initial 4D-CBCT images.
• Enhanced initial images lead to better motion modeling for MoCo reconstruction.
• The proposed method ultimately improves the quality of final 4D-CBCT images.