Anatomically-Aware Deep Learning for Medical Imaging: Hybrid GAN-Diffusion Framework for CT Image Synthesis from MRI and PET

Medical image translation is crucial for diagnostics, treatment planning, and radiation-free imaging. Traditional MRI-to-CT and PET-to-CT methods often struggle with structural inconsistencies and artifacts. This study introduces a CycleGAN-Diffusion hybrid model to enhance translation quality by combining CycleGAN’s unpaired image-to-image learning with diffusion models’ iterative denoising. Trained on MRI, PET, and CT datasets from Kaggle and TCIA, the model uses cycle consistency, adversarial, and SSIM loss for high-fidelity synthesis. Compared to RegGAN, CycleGAN, GcGAN, and Diffusion Models, our approach achieves superior SSIM and PSNR scores, reducing distortions and enhancing anatomical accuracy. This hybrid deep learning framework offers a promising step toward artifact-free, radiation-free synthetic CT imaging for clinical applications.