Mitigation of motion-induced artifacts in cone beam computed tomography using deep convolutional neural networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiation Oncology

Background:

• Cone Beam Computed Tomography (CBCT) is crucial for image-guided radiation therapy (IGRT), enabling accurate patient positioning.
• Motion artifacts in CBCT images can compromise treatment accuracy and adaptive capabilities.
• Deep learning offers a promising solution for mitigating these artifacts.

Purpose of the Study:

• To develop and evaluate a novel deep learning approach for reducing motion-induced artifacts in CBCT images.
• To improve the overall image quality of CBCT reconstructions.
• To integrate deep learning models as pre- and/or post-processing steps within the CBCT reconstruction pipeline.

Main Methods:

• Utilized deep convolutional neural networks, specifically refined U-net architectures, integrated with standard CBCT reconstruction methods (FDK, SART-TV).
• Trained neural networks end-to-end using a supervised learning setup with simulated motion data derived from 4D CT scans.
• Validated the approach using quantitative metrics (PSNR, SSIM) on test datasets and qualitative evaluation by clinical experts on real patient CBCT scans.

Main Results:

• The deep learning approach significantly reduced motion artifacts and improved CBCT image quality compared to existing methods.
• Achieved up to +6.3 dB and +0.19 improvements in PSNR and SSIM, respectively.
• Clinical evaluation showed up to 74% preference for motion artifact reduction using the proposed method over standard reconstruction.

Conclusions:

• Demonstrated the efficacy of deep neural networks, integrated as pre- and post-processing plugins, in enhancing CBCT image quality.
• Successfully reduced motion artifacts in CBCT images through end-to-end trained deep learning models.
• The study provides clinical evidence for the significant benefits of this deep learning approach in IGRT.