Leveraging GAN-based CBCT-to-CT translation models for enhanced image quality and accurate photon and proton dose calculation in adaptive radiotherapy

Abstract

PurposeTo develop a deep learning-based network for CBCT-to-CT translation. This approach aims to enhance CBCT image quality and enable more accurate photon and proton dose calculation in adaptive radiotherapy. MethodsThe proposed network is based on a simple generative adversarial network (GAN) and task-specific loss functions. The CBCT images were input into the network for synthetic CT (sCT) generation with regularization loss in the form of gradient loss (between CBCT and sCT) and style loss (between registered planning CT and sCT) to preserve the CBCT structural information and obtain accurate HU values through CT style transfer, respectively. CBCT and planning CT images collected from 58 patients with pelvic lesions were used for training. The model performance was quantitatively and qualitatively evaluated and compared to that of three other models (Res-G, CycleGAN and CUT). Quantitative analysis was performed to assess the noise reduction and dosimetric improvement achieved by the proposed network. The performance of the model was evaluated in an additional 6 pelvic patients. ResultsThe best visual effects on the test data were achieved using our model, which showed detailed information preservation and artifact removal ability. Quantitative values of 13.02 ± 4.40 HU, 56.59 ± 20.58 HU, 37.53 ± 3.06 dB and 93.43 ± 1.23% were obtained for the MAE, RSME, PSNR and SSIM, respectively. Our sCT also showed CT-like noise levels, with a mean global noise level (GNL) of 3.31 ± 0.32 HU. The dosimetric evaluation revealed a gamma pass rate higher than 99% under all criteria for photon treatment and high pass rates of 95.77 ± 1.21% (2%/2 mm) and 98.07 ± 0.45% (3%/3 mm)) for proton treatment. ConclusionWe introduce novel task-specific loss functions into a GAN-based network for generating high-quality sCT images. The model also has an excellent ability to remove artifacts and suppress noise. Dosimetric results show the potential for integration into the clinical ART workflow.