PAD Model Generates Heterogeneous PET Images from Organ Activity Maps

A pretrained domain-adapted diffusion (PAD) model has been created by researchers for synthesizing PET images conditioned on anatomy, utilizing uniform organ activity maps. This model incorporates a text-to-image decoder that is pretrained on natural images, along with an upstream conditioning encoder and a downstream adapter for the PET domain. The training process consists of two phases: initially learning the broad uptake distributions, followed by a refinement of local details. Uniform organ activity maps are derived from CT segmentations, where each organ is assigned its average uptake from corresponding PET images. The evaluation focuses on accuracy, noise levels, and radiomic analysis. This method seeks to address the challenges posed by traditional physics-based simulations, which are often resource-heavy and lack anatomical variability. The findings are detailed in arXiv:2605.20267.