CycleGAN is an image translation technique that can successfully suppress bones in dual-energy X-ray images. This study introduces two novel variations of the CycleGAN model, namely CycleGAN-Idempotent Loss (CGIL) and CGIL with additional layers (CGILAL). The PSNR, MS-SSIM, and VIF measures are used to compare the performance of CGIL and CGILAL models with CycleGAN and Pix2Pix GAN models. CGIL and CGILAL showed better performance than CycleGAN and Pix2Pix in terms of these metrics when applied to the dual-energy chest x-ray dataset. Here, these trained models are used to construct bone-suppressed images from standard X-rays in the CovidX dataset, where there are no equivalent bone-suppressed images available. The standard CXR images without bone suppression and bone-suppressed standard X-ray images are currently being used for the multi-class categorization of Covid, Pneumonia, and Normal images. A total of 16,965 images in each dataset are used for the study, with 5,655 images from each class. A total of 13,572 images are used for training, whereas 3,393 images are used for testing. The accuracy of standard CXR images without bone suppression is 90.07%. The accuracy of CGIL and CGILAL is better than that of CycleGAN and Pix2Pix, with CGIL achieving 95.14% accuracy and CGILAL achieving 95.76% accuracy, compared to CycleGAN’s accuracy of 94.49% and Pix2Pix’s accuracy of 94.22%. The findings indicate that models trained on bone-suppressed images exhibit superior accuracy in comparison to standard X-ray images without bone suppression.