Generative restoration for cardiac PET attenuation correction: a two-stage 3D DDIM framework optimizing fidelity and clinical controllability

This study aims to develop a two-stage 3D denoising diffusion implicit model (DDIM) framework for CT-free attenuation correction in cardiac PET imaging, enabling direct generation of attenuation-corrected (AC) images from non-attenuation-corrected (NAC) PET scans. The method is comprehensively validated using both 18FFDG PET and 13Nammonia cardiac PET datasets to demonstrate its clinical applicability across different perfusion and metabolic imaging protocols. The framework employs a two-stage approach: (1) a noise-to-image DDIM was first pretrained on all available AC images (i. e. , no need of paired NACs) to learn a diverse AC distributions, enabling the high-fidelity generation of AC images with varying appearances; (2) the pretrained model was fine-tuned with a limited set of paired NAC-AC images to form a conditional DDIM, ensuring anatomically aligned, controllable generation. The model architecture uses a 3D U-Net, trained on 224 paired NAC-AC and 396 unpaired AC images for 18FFDG, and 608 paired NAC-AC images and 885 unpaired AC images for 13Nammonia. Performance was evaluated through quantitative metrics (including NMAE, NRMSE, SSIM and PSNR) and visual assessment. The proposed two-stage DDIM framework achieved excellent agreement with clinical CT-based attenuation correction (CT-AC), demonstrating superior correlation (slope = 0. 78, \: R^2 = 0. 95 for 18FFDG; slope = 0. 99, \: R^2= 0. 91 for 13Nammonia) and lower errors compared to existing approaches. Ablation studies confirmed the benefits of both the two-stage training strategy and the incorporation of unpaired AC images, as evidenced by narrower confidence intervals in Bland-Altman analysis and reduced percentage errors. The two-stage 3D DDIM framework achieves performance comparable to clinical CT-AC while effectively leveraging unpaired data, demonstrating significant potential for robust cardiac PET attenuation correction.