<p>Cone-beam computed tomography (CBCT) images used for adaptive radiotherapy (ART) lead to the errors of modified treatment plans and the inaccuracy of the ART due to low quality, non-uniformity in Hounsfield unit (HU) values and morphological differences with planning CT (pCT) images. To overcome the issues, a Swin transformer-based CycleGAN (SCGAN) architecture was proposed for synthesizing CT images from CBCT images, and the perceptual loss was applied into network training for improving the performance of the SCGAN. The generator of the SCGAN had continuously-connected six residual Swin transformer blocks (RSTBs), which consisted of two Swin transformer layers (STLs) and a convolution layer. The STL was constructed using window-wise multi-head self-attention (W-MSA), shifted-window MSA (SW-MSA), layer normalization and multi-layer perceptron layers with skip connections. The perceptual loss was calculated through the VGG19 pre-trained with the ImageNet database, and the total loss function was defined by weighting adversarial, cycle consistency, identity and perceptual losses. The SCGAN model precisely generated synthetic CT (sCT) images, and the accuracy of the sCT images for the SCGAN model was higher than those for the Pix2Pix and CycleGAN models in terms of quantitative evaluation. Also, the performance of the SCGAN model was more improved by the perceptual loss. The SCGAN with the perceptual loss proposed in this study is able to improve the accuracy of the sCT image compared to the Pix2Pix and CycleGAN models, and the proposed model has a potential to deliver the sCT image closer to the GT image than the other models for the ART. The clinical availability of the proposed model would be validated through the investigation of dose distribution accuracy, the network training with well-matched image datasets, and the application of additional techniques for maximizing the SSIM of the sCT image.</p>