[Research progress on generative adversarial network-based CBCT-to-CT synthesis with multidimensional constraints]

In modern radiotherapy, computed tomography (CT) provides essential imaging information for treatment planning, while cone-beam computed tomography (CBCT) provides real-time anatomical information during treatment. However, scattering artifacts, noise, and CT value deviations limit the direct application of CBCT in adaptive radiotherapy. Generative adversarial networks (GANs) provide an effective approach for CBCT-to-CT cross-modal synthesis, but basic GANs may produce anatomical distortions and "anatomical hallucinations". In recent years, multidimensional constraint mechanisms have been increasingly introduced to improve the structural consistency and clinical reliability of synthetic CT (sCT). This review summarizes GAN-based CBCT-to-CT synthesis methods and classifies existing constraints into image-level, feature-level, and task-level constraints according to the level at which they are applied. The mechanisms, applicability, and limitations of different constraint strategies are compared. Furthermore, current challenges in structural fidelity, cross-domain mapping, and clinical task integration are discussed. Future directions, including generative framework optimization, multiscale feature modeling, and multitask integration, are also outlined. This review aims to provide a reference for the design and optimization of CBCT-to-CT synthesis models.