Author: Samuel Kadoury, Redha Touati 👨🔬
Affiliation: Polytechnique Montréal 🌍
Purpose:
Generating synthetic CT images from MR acquisitions for radiotherapy planning allows to integrate soft tissue contrast alongside density information stemming from CT, thus improving tumor and organ at risk delineations in treatment planning, while minimizing radiation exposure. This enables more accurate and personalized treatment strategies. Our work enhances MR-CT synthesis processes for head and neck radiotherapy planning by leveraging anatomical information within a diffusion model to improve tissue boundary and density accuracy. The goal is to improve the accuracy of tumor delineation from synthetic CTs, used to personalize treatment strategies.
Methods:
We used the Head and Neck Organ-at-Risk Segmentation (HaN-Seg) dataset, including paired CT and MRI-T1 scans from 56 head and neck cancer patients with CT-based organ-at-risk masks. The proposed structure-based generative diffusion model for synthesizing high-quality CT images from MR data, was implemented with a two-stage denoising diffusion probabilistic model (DDPM) framework. In the first stage, a pre-trained DDPM generates CT images from MRI-T1 inputs. The second stage refines these outputs by incorporating anatomical structures from the MRI data with the generated CT images using a secondary DDPM. Training was performed using a variational inference strategy.
Results:
The method's performance on the HaN-Seg dataset achieved a Structural Similarity Index of 0.85, a mean absolute error of 0.09, and a Peak Signal-to-Noise Ratio of 22.05. Additionally, the model exceled in tumor segmentation from synthetic sCT images, achieving a Probability Rand Index of 0.83, Dice score of 0.75, and Global Consistency Error of 0.16, underlining its precision in capturing tumor regions, improving segmentation accuracy compared to standard synthesis methods.
Conclusion:
Our structure-guided generative diffusion model demonstrated promising results in synthesizing CT images from MR acquisitions, yielding significant improvements compared to existing approaches. These results underscore its potential for advancing clinical applications, particularly in tumor imaging, diagnosis, and treatment planning.