Diffusion-Driven 3D CECT-to-NCECT Conversion for Head and Neck Imaging

Many medical applications, such as proton therapy and radiotherapy planning, require both Contrast Enhanced CT (CECT) for tumor and organ delineation and Non-Contrast Enhanced CT (NCECT) for accurate radiation dose calculation. The need for both scans can introduce dosimetric uncertainties due to tissue motion, particularly in thoracic and abdominal tumors. This study presents a diffusion model-based deep learning framework to generate NCECT directly from CECT, offering a solution that reduces the need for the NCECT scan and minimizes motion-related uncertainties. We trained and evaluated the diffusion model using 6400 2D CT slices from abdominal CECT and NCECT image pairs from 25 patients. The model's performance was evaluated using structural similarity index measure (SSIM), Mean Square Error (MSE), and qualitative visual assessments, focusing on the accuracy of NCECT image generation. The diffusion model achieved a SSIM score of 0.87 and a MSE of 0.005, significantly outperforming a previous Generative Adversarial Network (GAN) based approach. Qualitative assessments showed that the diffusion model generated images had superior tissue differentiation and anatomical detail, closely matching true NCECT images. These results highlighted the potential utility of the diffusion model to reduce the uncertainties associated with tissue motion between CT scans. The proposed method shows promise for clinical applications requiring high-quality and precise CT image synthesis.