Synthetic CT generation from MRI using 3D transformer-based denoising diffusion model

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiation Oncology

Background:

• Magnetic resonance imaging (MRI)-based synthetic computed tomography (sCT) simplifies radiation therapy planning.
• Eliminating CT simulation reduces patient radiation dose and setup uncertainty.

Purpose of the Study:

• To propose a transformer-based improved denoising diffusion probabilistic model (MC-IDDPM) for MRI-to-CT translation.
• To generate high-quality synthetic CT (sCT) from MRI to facilitate radiation treatment planning.

Main Methods:

• Developed MC-IDDPM using diffusion processes and a shifted-window transformer network (Swin-Vnet).
• Employed a forward process (adding noise) and a reverse process (denoising conditioned on MRI) to generate sCT.
• Evaluated the model on institutional brain and prostate datasets using quantitative metrics (MAE, PSNR, SSIM, NCC) and dosimetry analyses.

Main Results:

• MC-IDDPM achieved state-of-the-art quantitative results for brain sCT generation.
• Prostate sCT generation also showed strong performance with MAE 55.124 ± 9.414 HU.
• Statistically significant improvements (p < 0.05) were observed compared to competing networks, with dosimetry differences within ±0.34%.

Conclusions:

• Developed and validated a novel transformer-based DDPM for generating CT images from MRI.
• The model captures complex CT-MRI relationships, producing high-quality sCT efficiently.
• This approach can simplify radiation therapy planning, reduce patient time, and enhance treatment accuracy.