Equivariant conditional diffusion model for head and neck CT image synthesis from CBCT

Area of Science:

• Medical Imaging
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Cone-beam computed tomography (CBCT) is vital for image-guided radiotherapy (IGRT) due to its low cost and real-time visualization.
• CBCT images suffer from artifacts like inaccurate Hounsfield Unit (HU) values, compromising dose calculations and adaptive planning.
• Standard computed tomography (CT) offers superior image quality but cannot capture intra-treatment anatomical changes.

Purpose of the Study:

• Develop an accurate CBCT to CT synthesis method to address imaging quality gaps in adaptive radiotherapy.
• Improve the reliability of dose calculations and treatment planning in radiotherapy workflows.
• Mitigate artifacts in CBCT images for enhanced clinical decision-making.

Main Methods:

• Proposed EqDiff-CT, a novel diffusion-based conditional generative model for synthesizing high-quality CT images from CBCT.
• Employed a denoising diffusion probabilistic model (DDPM) to learn latent representations for image reconstruction.
• Utilized a group equivariant conditional U-Net backbone with e2cnn steerable layers to enforce rotational equivariance and preserve structural details.

Main Results:

• EqDiff-CT demonstrated significant improvements in structural fidelity, HU accuracy, and quantitative metrics compared to CycleGAN and DDPM.
• Visual analysis confirmed enhanced image recovery, sharper soft tissue boundaries, and more realistic bone reconstructions.
• The model was trained and validated on the multi-site SynthRAD2025 dataset.

Conclusions:

• The proposed diffusion model provides a robust and generalizable framework for improving CBCT image quality.
• EqDiff-CT enhances image quality and clinical confidence in CBCT-guided treatment planning and dose calculations.
• This advancement supports more accurate and reliable adaptive radiotherapy.