New PAD Diffusion Model Generates Lifelike PET Scans from Organ Activity Maps

Researchers have introduced PAD (Pretrained Domain-Adapted Diffusion Model), a new approach for generating clinically realistic PET images from simple uniform organ activity maps. Conventional physics-based PET simulation is computationally expensive, limited in anatomical variety, and fails to capture heterogeneous tracer uptake. PAD overcomes these limitations by leveraging a natural-image pretrained text-to-image decoder as its backbone, augmented with an upstream conditioning encoder and a downstream PET-domain adapter. A two-phase training strategy first learns coarse uptake distributions from segmentations, then refines local image details using paired PET data. This enables the model to produce heterogeneous PET images that match real scans in noise texture, quantitative accuracy, and anatomical fidelity.

In rigorous testing, PAD-generated images showed concordance correlation coefficients above 0.92 between organ mean standardized uptake values (SUVs) and the assigned activities. Radiomic analysis confirmed texture characteristics similar to target images, and automatic tumor segmentation performance was comparable. In a two-alternative forced-choice human observer study, four readers achieved approximately 50% accuracy, meaning they could not visually distinguish synthetic from real PET scans. The model also demonstrated compatibility with XCAT digital phantom anatomies, generating realistic images from non-clinical priors. PAD provides a scalable, diffusion-based framework for data augmentation in medical AI training, virtual imaging trials, and quantitative workflow development—reducing reliance on costly patient data.