Boundary information-guided adversarial diffusion model for efficient unsupervised synthetic CT generation.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Magnetic Resonance Imaging (MRI) lacks electron density information (Hounsfield Units), limiting its use in radiotherapy (RT).
• Synthetic CT (sCT) from MRI simplifies RT planning and improves accuracy by eliminating CT simulation, radiation dose, and registration errors.
• Existing unsupervised methods like CycleGAN struggle with structural consistency in sCT synthesis.

Purpose of the Study:

• To develop RadADM, a novel unsupervised boundary information-guided adversarial diffusion model.
• To enhance unpaired MR-to-CT translation for MR-only RT applications.

Main Methods:

• RadADM incorporates boundary mask information to guide feature learning and anatomy compensation during sCT generation.
• A cycle-consistent module with adversarial projections and coupled diffusion/non-diffusion architecture facilitates training on unpaired datasets.
• Performance was validated against state-of-the-art methods including CycleGAN, CycleSlimulationGAN, CUT, F-LseSim, and SynDiff.

Main Results:

• RadADM outperformed comparative methods in generating high-quality sCT from pelvic MRI datasets.
• The model achieved superior capture of local features with lower Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).
• Quantitative metrics showed high similarity: PSNR 24.70 ± 0.52, SSIM 0.8673 ± 0.01 overall; soft-tissue: PSNR 33.99 ± 1.09, SSIM 0.931 ± 0.01; bone: PSNR 35.79 ± 0.87, SSIM 0.993 ± 0.04.

Conclusions:

• RadADM effectively synthesizes anatomically accurate sCT, demonstrating robustness on pelvic datasets.
• The approach offers a promising direction for clinical MR-only adaptive radiotherapy, particularly for pelvic cancer treatment.