HDDPM model cuts PET scan radiation dose by 99% while preserving image quality

Positron emission tomography (PET) imaging faces a fundamental trade-off between diagnostic quality and radiation dose. Low-count PET scans introduce non-Gaussian, spatially dependent noise that scales with local activity—a complexity that standard denoising diffusion probabilistic models (DDPMs) fail to capture because they use isotropic, homoscedastic Gaussian noise. To address this, Raymond Confidence and Udunna C. Anazodo propose HDDPM, a heteroscedastic residual diffusion model that makes the forward corruption process intensity-aware. They design a fixed, Poisson-based variance module that generates voxel-wise noise maps, placing stronger perturbations on low-activity regions while the network predicts the low-to-standard-count residual under explicit dose-fraction conditioning.

Evaluated across three different PET scanners using both internal and external datasets at simulated dose levels from 1% to 50%, HDDPM and isotropic DDPM showed comparable overall image quality. However, HDDPM stood out in the lowest-dose (1%) external scans, demonstrating high reliability and significantly reducing measurement errors in both high- and low-activity regions. These results confirm that heteroscedastic noising with HDDPM provides a physically motivated inductive bias for quantitative low-count PET recovery, potentially enabling safer, lower-radiation brain scans without sacrificing diagnostic accuracy.