Structural Guidance in Stacked Generative Diffusion Model: Synthesizing Head and Neck CT from MRI in Radiotherapy...

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Head and neck radiotherapy relies on both MRI for soft tissue detail and CT for radiation planning.
• Current methods necessitate CT scans, increasing patient radiation exposure.
• Synthesizing CT data from MRI can reduce exposure and enable adaptive radiotherapy planning.

Purpose of the Study:

• To propose a novel generative diffusion model for synthesizing CT images from MRI data in head and neck radiotherapy.
• To improve the accuracy and utility of synthetic CT (sCT) images for dosimetry and adaptive re-planning.

Main Methods:

• Utilized a stacked denoising diffusion probabilistic model (DDPM) framework with two stages: a structure image generator and a contextual image generator.
• Incorporated structural guidance from MRI into the synthesis process using an augmented multi-channel input.
• Employed a variational inference training approach combining variational lower bound loss and mean absolute error loss.

Main Results:

• The model achieved superior performance compared to existing MR-to-CT generative models on the HaN-Seg dataset.
• Quantitative metrics included a multiscale-SSIM of 0.85 ± 0.08, MAE of 0.09 ± 0.06, and PSNR of 22.05 ± 1.83.
• High scores in segmentation accuracy (PRI: 0.83 ± 0.04, Dice: 0.75 ± 0.07) were observed for the tumor area in synthetic CT images.

Conclusions:

• The proposed stacked diffusion model effectively synthesizes CT images from MRI, offering a promising alternative to conventional CT scans.
• This method has the potential to significantly minimize radiation exposure for patients undergoing head and neck radiotherapy.
• The generated sCT images provide sufficient accuracy for dosimetry and adaptive re-planning, enhancing radiotherapy workflows.