Disentangling structure from modality: An attention-based framework for CBCT-to-CT synthesis

Cone-beam CT (CBCT) is essential for daily image guidance in radiotherapy; nevertheless, image degradation due to scatter and noise limits its quantitative application. Current deep learning techniques frequently compromise anatomical accuracy by indiscriminately altering all latent characteristics. We present an attention-guided generative adversarial network (AGD-GAN) that maintains anatomical structures by learning to provide a pixel-level spatial mask. This mask, based on a ResNet-50 encoder and concurrent attention modules, distinctly differentiates modality-invariant anatomy from modality-specific appearance. A compact transformation network maps solely the appearance features to the target domain, while the isolated anatomical features are transmitted unaltered to a StyleGAN2-based decoder. This distinction is maintained by a latent feature-consistency loss that offers direct structural oversight. Assessed using the public SynthRAD2023 benchmark (five-fold cross-validation, 1080 paired volumes), our approach attained a mean absolute HU error of 42.72, surpassing robust GAN and diffusion baselines. The robustness was validated in an external retrospective cohort of 25 patients. Our methodology provides a reliable and efficient solution for generating high-fidelity synthetic CT by developing an interpretable mask of the preserved anatomy, thereby directly enhancing adaptive radiotherapy operations.