Advancing PET image reconstruction: from MAP to generative AI

Brief tutorial and history of conventional PET image reconstruction, including deep learning based approaches, and finishing with methods using a diffusion model within image reconstruction
ABSTRACT
PET image reconstruction has been in ongoing development over many decades, due to the need for improved image quality when reconstructing from raw PET data with limited counts and limited spatial resolution. Major advances in reconstruction have come from improving the model of the data (modelling the imaging physics as well as the noise), and improving the model of the reconstructed images (choice of basis functions, and/or choice of regularisation to compensate for noise). Up until recently, therefore, maximum a posteriori (MAP) image reconstruction, which combines all of these advances, represented the state of the art in PET reconstruction, by combining improved data models with more advanced image models (using, for example, anatomically-informed prior information to reduce image noise and improve spatial resolution). Nonetheless, MAP reconstruction methods still rely either on relatively simple prior information (e.g. the relative difference prior) or potentially overly strong prior information (such as anatomical guidance). This talk will start from these foundations and then cover recent progress in the use of deep learning to provide even more powerful modelling for PET image reconstruction. The use of supervised deep learning through to the use of generative AI methods for reconstruction will be covered, where in the case of generative AI, no longer is a single reconstructed image obtained, but instead multiple reconstructions are generated, corresponding to samples from a learned posterior distribution. Finally, if time permits, ultra-fast self-supervised MAP image reconstruction methods will also be discussed.