Self-contained deep learning-based boosting of 4D cone-beam CT reconstruction

Area of Science:

• Medical Imaging
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Four-dimensional cone-beam computed tomography (4D CBCT) is crucial for radiotherapy (RT) of moving targets like lung and liver.
• Interfraction motion variability necessitates high-quality imaging for accurate treatment.
• Current 4D CBCT suffers from sparse view artifacts, limiting motion quantification.

Purpose of the Study:

• To introduce a deep learning framework to enhance 4D CBCT image quality.
• To improve the accuracy of motion quantification in radiotherapy.
• To provide a versatile solution compatible with existing CBCT reconstruction and clinical workflows.

Main Methods:

• A deep learning framework utilizing a residual dense network (RDN) was developed.
• The RDN learns the relationship between sparse-view and full-view CBCT images to remove artifacts.
• The approach does not require patient-specific prior knowledge, making it self-contained.

Main Results:

• The framework consistently reduced streak artifacts in 4D CBCT phase images across different datasets and reconstruction methods.
• Root mean squared error was reduced by approximately 50%, and normalized cross-correlation improved by up to 10% compared to ground truth.
• Enhanced image quality led to more plausible motion fields estimated by deformable image registration (DIR).

Conclusions:

• The proposed framework significantly boosts 4D CBCT image quality.
• It improves deformable image registration (DIR) and motion field consistency.
• This facilitates motion information extraction from artifact-laden images, a key challenge in RT integration.