Improving synthetic CT generation with enhanced registration

Motivation: Synthetic CT (sCT) generation from MRI can provide mineralized tissue data without ionizing radiation. Voxel to voxel correspondence between MRI and CT for supervised training is one of the challenges in the generation of sCT. Goal(s): This study aims to evaluate the impact of registration techniques on sCT bone generation quality. Approach: Two registration methods, image-based and bone-centered, were applied to pre-process MRI-CT data before training 3D U-Net models for sCT generation. Model performance was assessed using MAE and PSNR. Results: Bone-centered registration significantly improved sCT accuracy, with lower MAE and higher PSNR. Impact: Bone-centered registration enhances synthetic CT generation for bone imaging, potentially advancing musculoskeletal imaging applications focused on bone pathology.