CT-Less Attenuation Correction Using Multiview Ensemble Conditional Diffusion Model on High-Resolution Uncorrected PET Images

Positron emission tomography requires precise quantification to ensure reliable diagnostic outcomes and treatment monitoring. The attenuation of emitted photons in tissues before detector capture poses a significant challenge. Without appropriate correction methods, attenuation can introduce quantitative bias, making it difficult to differentiate benign from malignant conditions, and potentially leading to misdiagnosis. Traditional attenuation correction relies on co-computed tomography (CT) scans, providing anatomical information to calculate signal loss. Unfortunately, this approach exposes patients to additional radiation, is prone to misalignment between scans, and requires expensive hardware. Recent innovations in deep learning offer an alternative solution through pseudo-CT generation. Our research demonstrates that Conditional Denoising Diffusion Probabilistic Models (DDPMs) significantly outperform previous state-of-the-art UNet approaches. By utilizing all three orthogonal views from non-attenuation-corrected PET images, the DDPM approach combined with ensemble voting generates higher quality pseudo-CT images with reduced artifacts and improved slice-to-slice consistency. Results from a study of 159 head scans acquired on the Siemens Biograph Vision PET/CT scanner show both qualitative and quantitative improvements in pseudo-CT generation using this diffusion model approach with an average absolute error of $(0.34 \pm 0.35) \%$ for a typical PET slice compared to the PET image reconstructed with the CT.