CBCT-to-CT synthesis using a hybrid U-Net diffusion model based on transformers and information bottleneck theory

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiation Therapy
• Image Processing

Background:

• Cone-beam computed tomography (CBCT) is crucial for image-guided radiation therapy (IGRT) but suffers from artifacts and noise.
• Improving CBCT image quality to CT-like standards is essential for precise radiation treatment.

Purpose of the Study:

• To develop a novel deep learning model for synthesizing high-quality CT-like images from CBCT scans.
• To enhance the precision and clinical applicability of CBCT in IGRT.

Main Methods:

• Proposed a hybrid U-Net diffusion model (HUDiff) integrating Vision Transformer (ViT) for global feature extraction.
• Incorporated a variational information bottleneck for efficient feature compression and filtering.
• Introduced a dynamic modulation factor to optimize the denoising process.

Main Results:

• The HUDiff model demonstrated superior performance in improving CBCT image quality compared to state-of-the-art methods.
• Experiments on Brain and Head & Neck datasets validated the model's robustness and clinical applicability.
• Synthesized images achieved CT-like quality, reducing artifacts and noise.

Conclusions:

• The HUDiff model effectively enhances CBCT image quality, addressing limitations in current IGRT practices.
• This advancement has the potential to significantly aid clinicians in developing more accurate radiation therapy plans.
• The model shows promise for future clinical integration to improve patient outcomes.