Online adaptive radiotherapy (oART) is becoming more widely used in clinical practice worldwide due to its ability to re-optimize treatment plans to account for daily patient-specific treatment variations (1, 2). The first clinical workflow programs for online ART were established using magnetic resonance guided radiotherapy (MRgRT) (3, 4) and Ethos^® Therapy was the first oART solution commercialized by Varian Medical System using cone beam computed tomography (CBCT) images (5).

The use of CBCT for dose calculation holds significant appeal when it comes to dose reporting and monitoring, particularly in the context of dose-guided adaptive radiotherapy. Although CBCT has revolutionized image-guided radiotherapy (IGRT) and is extensively used for daily patient setup (6), its use for plan adaptation is still limited. The main reasons are the artefacts related with scatter and the reconstruction algorithms, which reduce image quality and CBCT-based dose calculation accuracy (7–9). Several solutions have been proposed using deformable image registration (10–12), scatter correction (13, 14), or rescaling of Hounsfield unit intensities (15).

With the development of artificial intelligence, studies have focused on deep learning approaches (16–19). Kia et al. (20) was the first team to create a so-called synthetic-CT from a CBCT. They developed a deep convolutional neural network method for improving CBCT image quality for prostate cancer patients. Since then, the number of investigations have grown, resulting in many different architectures. In the recent years, the generative adversarial network (GAN) is one of the most developed architectures due to the anatomical preservation and the use of unpaired data sets. Xue et al. (21) generated sCT from head and neck CBCT using the CycleGAN, Pix2Pix and U-Net models and the performance of CycleGAN was proved to be best among three models in terms of image quality and dose accuracy. Likewise, Pang et al. (22) compared four deep learning architectures (cycleGAN, Unet, Unet + cycleGAN, conditional GAN) for generating nasopharynx CBCT and showed the superiority of the conditional GAN model in an adaptive proton treatment planning application. Maspero et al. (23) also demonstrated the feasibility to use a single cycle-GAN architecture for different anatomical sites. However, few studies investigated the model training using multi-center datasets (24) as proposed for MR-to-CT synthesis (25, 26). This study introduces and evaluates an artificial intelligence–based method for synthetic CT (sCT) generation in a pre-clinical validation setting to address these challenges, with the aim of fully harnessing the potential of CBCT for adaptive radiotherapy in patients with pelvic, head and neck, lung, and breast cancers.

Training of the synthetic CT model required approximately two weeks. The images were resampled to an isotropic resolution of 1 mm³ and cropped or padded to a fixed size of 16 × 512 × 512 voxels (Z, Y, X). Model convergence was achieved after approximately 200 training epochs using NVIDIA RTX 3090 GPUs. Using two GPUs, the generation of a sCT image took about 30 seconds, with an additional 20 seconds for augmentation, resulting in a total generation time of roughly 50 seconds per case. Generation time may vary depending on hardware configuration and imaging parameters.

Figure 1 presents representative examples of sCTs generated by the AI-based model, illustrating the comparison of Hounsfield Units with the corresponding deformably registered CT images for the male pelvis, thorax, head and neck, and breast regions. The image quality of sCTs was markedly improved compared with CBCT, primarily due to reduced scatter artifacts and smoother the HU distributions.

The Table 2 summarizes the quantitative image similarity metrics for each anatomical site. Overall, the MAE decreased from 122.95 ± 50.07 to 23.65 ± 10.09, the SSIM increased from 0.78 ± 0.12 to 0.97 ± 0.03 and the PSNR increased from 35.01 ± 7.24 to 44.35 ± 7.07 when comparing CBCT and sCT images. Despite variations in CBCT image quality between Varian and Elekta acquisitions, only minor differences were observed in the corresponding sCT reconstructions (Table 3).

Figure 2 shows the relative dose difference for the PTV and the organs at risk between plans calculated using the synthetic CT and the warped CT for pelvis, thorax, head and neck and breast localizations. The median relative dose differences for the PTV were lower than 0.5% for each dose metric (D_98%, D_95%, D_50%, and D_2%). The maximum reported value was 2.95% for D_98%, equivalent to 1.05Gy for the thorax. Concerning the organs at risk, the median relative dose difference were less than 0.5%, regardless the dose metrics and localizations assessed. The maximal difference was -4.46% for V30Gy of the heart, corresponding to 0.10Gy. The largest relative dose difference for PTV and heart are from the same patient illustrated in Figure 3.

The median global gamma analysis using two tolerance criteria (3%/3 mm and 2%/2 mm) with a 10% dose threshold demonstrated strong agreement between the dose distributions calculated on the synthetic CT and the warped CT (Table 4). The lowest gamma passing rate was observed for the thorax case, with a value of 96.6% for the 2%/2 mm criterion. This corresponded to the same patient presented earlier (Figure 3).

This study introduced 3D cycle GAN models for generating synthetic-CTs from CBCT images in cases of pelvic, head and neck, lung, and breast cancer. The key strengths of this work include its multi-centric approach, incorporating CBCT images from various linac manufacturers with differing characteristics, and the extensive patient dataset used for model training, ensuring robust synthetic-CT generation. Additionally, the cycle GAN architecture’s ability to function without paired images is particularly beneficial for areas like the pelvis, where anatomical variations are frequent.

Regarding the qualitative evaluation, the results obtained in this study are satisfying, showing a good agreement between the synthetic-CT and the warped CT. In addition, this research demonstrated that AI-driven synthetic CT generation from CBCT is clinically accurate across multiple anatomical sites. In terms of image accuracy, the MAE was 23, 65 ± 10, 09 for all confounded localizations. Dosimetric assessment showed median dose differences within 0.5% and median gamma pass rates of 100% for both the 3%/3mm and 2%/2mm criteria. Our results are consistent with recent literature that extends image quality analysis to include additional dose assessment (28–32). For example, O’Hara et al. (30) evaluated the image and dose accuracy of AI-guided sCTs generated from head and neck CBCT images. They reported a MAE of 79.4 HU, a gamma pass rate of 99.8% (2%/2mm, 20% dose threshold) and a generation time of 112 seconds. However, their dosimetric evaluation was limited to the CBCT field of view, which does not encompass the entire patient volume, potentially limiting its effectiveness in an adaptive workflow. Xie et al. (31) developed a technique similar to ours for handling truncated CBCTs in breast cancer cases and demonstrated excellent dose calculation accuracy with gamma pass rates of 98.98 ± 0.64% and 99.69 ± 0.22% for 2%/2mm and 3%/3mm criteria, respectively.

A primary limitation of this study is that the planning CT was used as the ‘ground truth’ to evaluate the quality of the sCTs generated from CBCT data. Despite using a wrapped CT to reduce discrepancies, anatomical variations may still be introduced by the time interval between the planning CT and the CBCT acquisition, such as differences in organ filling, air cavity fluctuations, or tumour regression, that can affect the agreement between the two image sets. Consequently, as illustrated in Figure 3, some of the observed deviations may be due to imperfections in the reference images rather than to limitations in the sCT generation process itself. Furthermore, although differences in CBCT image quality were observed between the Varian and Elekta systems, these did not appear to cause any noticeable deterioration in the corresponding sCTs. However, the relatively small number of test cases per anatomical site and per manufacturer reduces the strength and generalisability of these findings. Therefore, expanding the patient cohort in future studies will be essential to improve the statistical robustness of the conclusions. Finally, the sCT generation framework was evaluated in a pre-clinical validation setting and was not clinically deployed at the time of the study. Consequently, the results reflect a controlled research environment and may not fully capture the constraints and variability encountered in routine clinical workflows. While deformable image registration between planning CT and CBCT is a mature and commercially available approach for adaptive radiotherapy (Ethos, Raysearch, RadFormation, SeeTreat, …), deep learning–based sCT generation remains an emerging technique with limited clinical implementation (30). Therefore, the reproducibility and generalizability of the present findings will require confirmation through prospective clinical studies, including assessment of robustness, workflow integration, and regulatory considerations, before widespread clinical adoption can be considered.

The growing number of sCT generators developed using different deep learning methods presents a challenge too, as these networks are often trained on different datasets and anatomies, and evaluated using different metrics. This inconsistency makes it difficult to perform a standardized methodological comparison, which in turn hinders the ability to identify the optimal network design choices for clinical sCT tools. In this context, the SynthRAD2023 challenge (covering brain and pelvic patients), followed by SynthRAD2025 (covering head-and-neck, thorax, and abdomen), was designed to compare and benchmark synthetic CT generation methods using multi-center ground truth data (33–35). The SynthRAD2023 findings indicated no significant correlation between image similarity metrics and dosimetric accuracy, thereby highlighting the necessity of comprehensive dose evaluation when assessing the clinical viability of sCT in radiotherapy workflows (33). Furthermore, the absence of standardized criteria for assessing and reporting sCT quality has been recognized as a key factor impeding the clinical adoption of new methods (36). Developing a robust quality assurance process will be a future challenge to ensure the safe and reliable integration of deep learning technologies into clinical workflows (37).

This retrospective multi-center study has shown the potential of synthetic CT across various anatomical sites. This AI-based tool is clinically acceptable, allows for significantly faster image conversion, and may enhance the implementation of adaptive radiotherapy treatments.