Background and objective: CT-based attenuation correction (CT-AC) is commonly used in cardiac PET but introduces additional radiation exposure, which is particularly problematic for pediatrics or repeat scans. This study aims to generate attenuation-corrected cardiac PET images from non-attenuation-corrected (NAC) PET input without CT using a conditional denoising diffusion implicit model (cDDIM).

Methods: The patient cohort included 60 rest-only and 14 rest-stress scans from Center 1, and 30 rest-stress scans from Center 2. All subjects underwent ¹³N-ammonia cardiac PET/CT scans, generating paired 3D NAC and CT-AC PET images. The middle slice and two adjacent axial slices were used as inputs to mitigate axial artifacts, forming a 2.5D cDDIM framework. Two versions were proposed: one with data logarithmic linear normalization (cDDIM_LLN) and another with linear normalization (cDDIM_LN). 2.5D and 3D generative adversarial network (GAN)-based AC methods were implemented for comparison. Model performances were evaluated using normalized mean square error (NMSE) and segment-wise absolute percent error (APE).

Results: Visual analysis indicated that cDDIM_LLN and cDDIM_LN outperformed GAN-based AC methods for both rest and stress cardiac PET images. cDDIM_LLN achieved significantly lower NMSE than cDDIM_LN in Center 1 (1.87 ± 1.25 % vs. 2.77 ± 2.37 %, p < 0.001) and Center 2 (4.63 ± 3.71 % vs. 5.67 ± 7.88 %, p < 0.001). cDDIM_LLN also showed smaller APE than other methods for both centers (Center 1: 5.78 ± 1.43 %, p < 0.001; Center 2: 9.10 ± 5.59 %, p < 0.001). cDDIM_LN demonstrated lower APE than GAN-based AC methods across both centers.

Conclusions: cDDIM-based AC methods synthesized tracer distributions highly similar to clinical CT-AC. Among the evaluated methods, cDDIM_LLN demonstrated the best performance. Overall, cDDIM-based AC outperformed traditional GAN-based methods.