Abstract
Cone-beam computed tomography (CBCT) scans are widely used for real time monitoring and patient positioning corrections in image-guided radiation therapy (IGRT), enhancing the precision of radiation treatment. However, compared to high-quality computed tomography (CT) images, CBCT images suffer from severe artifacts and noise, which significantly hinder their application in IGRT. Therefore, synthesizing CBCT images into CT-like quality has become a critical necessity. In this study, we propose a hybrid U-Net diffusion model (HUDiff) based on Vision Transformer (ViT) and the information bottleneck theory to improve CBCT image quality. First, to address the limitations of the original U-Net in diffusion models, which primarily retain and transfer only local feature information, we introduce a ViT-based U-Net framework. By leveraging the self-attention mechanism, our model automatically focuses on different regions of the image during generation, aiming to better capture global features. Second, we incorporate a variational information bottleneck module at the base of the U-Net. This module filters out redundant and irrelevant information while compressing essential input data, thereby enabling more efficient summarization and better feature extraction. Finally, a dynamic modulation factor is introduced to balance the contributions of the main network and skip connections, optimizing the reverse denoising process in the diffusion model. We conducted extensive experiments on private Brain and Head & Neck datasets. The results, evaluated from multiple perspectives, demonstrate that our model outperforms state-of-the-art methods, validating its clinical applicability and robustness. In future clinical practice, our model has the potential to assist clinicians in formulating more precise radiation therapy plans.