Generative structure-guided denoising diffusion probabilistic model for MR to CT image synthesis in radiotherapy planning.
Authors
Affiliations (2)
Affiliations (2)
- Ecole Polytechnique de Montreal, 2500 chemin Polytechnique, Montreal, H3C3A7, Canada.
- CRCHUM, Montréal, Québec, Canada, 1000 R. Saint-Denis, Montréal, QC, Montreal, Quebec, H2X 0C1, Canada.
Abstract
Radiotherapy (RT) planning uses magnetic resonance (MR) imaging for soft tissue information, while computerized tomography (CT) imaging is necessary for dose calculation. Synthetic CTs generated from MR images streamline the clinical workflow by reducing the need for an actual CT acquisition while minimizing radiation exposure, enabling adaptive RT workflows. 
Appraoch: We propose a radiotherapy planning model combining stable structure-guided denoising diffusion probabilistic model (DDPM) with MRI conditioning, for generating synthetic CT (sCT). The model combines a hybrid network with a dynamic guidance mechanism, conditioned on MR, to direct the reverse diffusion process and preserve structural details. The model integrates both a denoising and a refinement network, while the guidance strength adapts at each timestep based on the noise level. The training process includes a weighted structure-guided loss function that combines the DDPM noise loss with a dynamic guidance regularizer term for sCT image generation, with a refinement network. 
Main results: Evaluated on three paired MR/CT datasets (head and neck, brain), our model outperforms recent MR to CT synthesis models, showing the best performance in image quality based on the Mean Absoluate Error (MAE) in Hounsfield units, as well as improved Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM) and Gradient Magnitude Similarity Deviation (GMSD) compared to other DDPM and GAN approaches. When comparing tumor segmentation quality, it achieves the best results in Dice score (0.76 ± 0.04) for segmented tumor regions in the generated sCT images. Finally, it achieves the best Mean Square Error (MSE) of 10.1 ± 2.6 Gray in the head and neck cohort, when evaluating dose planning on sCT, with a Conformity Index (CI) of 0.61 ± 0.02 and a Homogeneity Index (HI) of 0.10 ± 0.02. 
Significance: Performance shows promise in integration for radiotherapy in head and neck, as well as in brain cancers.