Fidelity-oriented pelvic CBCT-to-CT synthesis using an anatomical information-attention diffusion model

Cone-beam computed tomography (CBCT) is widely used in image-guided radiotherapy and adaptive radiotherapy, but its direct quantitative use is limited by scatter artifacts, low soft-tissue contrast, and inaccurate Hounsfield unit (HU) values. Diffusion models provide stable image generation capabilities for CBCT-to-CT synthesis. However, in medical image synthesis tasks where patient-specific anatomical fidelity is more important than generation diversity, the stochastic sampling capability of diffusion models may conflict with task requirements. We propose a fidelity-oriented Anatomical Information-Attention Diffusion Model (AIADM) for pelvic CBCT-to-CT synthesis to generate synthetic CT (sCT) images. Unlike conventional diffusion models that primarily predict noise during reverse denoising, AIADM directly predicts the image state at the preceding diffusion time step, aiming to reduce unnecessary generation variation. A Fourier-based anatomical information fusion module integrates frequency-domain features of the current diffusion state with CBCT structural features, while a Vision Transformer attention module captures global pelvic structural dependencies. The diffusion process loss and anatomical information consistency loss constrain the intensity residuals of stepwise prediction and the target edge responses, respectively. Experiments on an internal pelvic dataset comprising 86 patients show that AIADM outperforms the GAN-based, VAE-based, and diffusion-based methods compared in this study across the four evaluated metrics. Using the registered reference pCT as the evaluation reference, AIADM achieves an MAE of 14.25 ± 1.05 HU, an RMSE of 26.42 ± 5.70 HU, a PSNR of 43.87 ± 1.89 dB, and an SSIM of 0.96 ± 0.01. Visual comparisons and HU profiles show that the generated results are closer to the reference pCT, with fewer visible streak artifacts in multiplanar and three-dimensional reconstruction views. AIADM introduces image-domain supervision and anatomical feature fusion into a conditional diffusion framework for pelvic CBCT-to-CT synthesis. The proposed framework reduces intensity errors relative to the registered pCT, increases image similarity, and improves the visual continuity of reconstruction views, providing a basis for subsequent image synthesis and dosimetric evaluation.