Boundary information‐guided adversarial diffusion model for efficient unsupervised synthetic CT generation
Changfei Gong, Junming Jian, Yuling Huang, Mingming Luo, Shenggou Ding, Xingxing Yuan, Xiaoping Wang, Yun Zhang · Medical Physics · 2025
BACKGROUND: The absence of tissue electron density information derived from greyscale Hounsfield units (HUs) in magnetic resonance imaging (MRI) limits its further clinical application in radiotherapy (RT). The use of synthetic computed tomography (sCT) with MRI simplifies RT treatment and improves positioning accuracy by eliminating the need for computed tomography (CT) simulation with radiation dose and error-prone image registration. Although CycleGAN and its variants can obtain verisimilar sCT through unsupervised learning, ensuring perfect structural consistency of the synthesized images in this approach remains challenging, and thus limiting the quality and diversity of the images synthesized for a given application. PURPOSE: The purpose of this work is to develop a novel unsupervised boundary information-guided adversarial diffusion model, called RadADM, with the aim of enhancing performance in regard to unpaired MR-to-CT translation for MR-only RT. METHODS: In order to explicitly guide the feature learning of the proposed RadADM model, the boundary mask information is incorporated as guidance for anatomy compensation during sCT generation from simulated MR images. In addition, a cycle-consistent module incorporates adversarial projections featuring coupled diffusive and non-diffusive architecture is used to facilitate training on unpaired MR-CT datasets, enabling accurate and efficient translation between the source and target domain images. To validate the performance of the proposed model, we conducted a comprehensive quantitative and qualitative comparison of RadADM with other state-of-the-art methods, including CycleGAN, CycleSlimulationGAN, CUT, Fixed Learned Self-Similarity (F-LseSim), and SynDiff. RESULTS: We evaluated and demonstrated that RadADM outperforms other comparative approaches for high-quality sCT generation on pelvic MRI datasets, captures high-quality local features, and achieves smaller errors with mean absolute error (MAE): 62.95 ± 23.15 and root mean square error (RMSE): 135.46 ± 23.89 and higher similarities with peak signal-to-noise ratio (PSNR): 24.70 ± 0.52, structural similarity index (SSIM): 0.8673 ± 0.01. For the region of soft-tissue, the PSNR and SSIM were 33.99 ± 1.09 and 0.931 ± 0.01, and for the region of bone, the PSNR and SSIM were 35.79 ± 0.87 and 0.993 ± 0.04. CONCLUSIONS: Extensive experiments on pelvic datasets demonstrate the effectiveness and robustness of our proposed RadADM in terms of enabling synthesizing sCT at the anatomical level. Our approach is found to offer a valuable and promising direction for clinical MR-only adaptive radiotherapy for pelvic cancer.