A new AI model improves brain PET scan quality at radiation doses as low as 1% of standard — by teaching the denoising process to respect the physics of the scanner.
PET imaging forces a tradeoff: lower radiation dose means noisier images, and that noise is not simple or uniform. It scales with local tissue activity, shifts depending on where you look in the image, and is shaped by the reconstruction algorithms the scanner runs internally. Standard denoising diffusion models ignore all of that — they add the same flat, Gaussian noise everywhere during training, which does not resemble what a real PET scanner produces. The new model, called HDDPM, fixes this by building a Poisson-based noise module that generates voxel-level noise maps. Low-activity regions get hit harder; high-activity regions less so — matching the actual physics.
The practical payoff shows up at the extremes. Across three different scanners and both internal and external datasets, HDDPM and a standard diffusion model performed comparably at moderate dose reductions. At 1% dose — the hardest case — HDDPM pulled ahead, cutting measurement errors in both high- and low-activity regions. For quantitative PET, where clinicians use the numbers to track disease or assess treatment response, those errors are not just cosmetic.
The result is a reminder that domain knowledge still earns its keep. General-purpose generative models keep getting better, but building the right physical prior into the noise schedule — rather than patching it in after the fact — appears to matter most when conditions are worst.