A research team has built a neural network that generates Tau-PET brain scans from standard MRI images, skipping the radioactive tracers and specialized scanners that make PET scans hard to scale.
Tau-PET imaging is a key tool for staging Alzheimer's disease — it tracks the spread of tau protein tangles through the brain. But PET scanners are expensive, tracer supply is limited, and the radiation dose adds a barrier to routine use. SFL-Net, introduced in a new paper, takes T1-weighted and FLAIR MRI scans as input and synthesizes a predicted Tau-PET image. The model splits its internal representation into shared, T1-specific, FLAIR-specific, and complementary pathways, then uses structural conditioning in the latent space rather than the skip connections common in UNet-style architectures. Researchers trained and validated it on 605 and 83 subjects respectively, drawn from two established datasets, ADNI-3 and OASIS-3.
The practical upside goes beyond image quality. Because SFL-Net factorizes its sources, clinicians can audit which input — T1 or FLAIR — drove a given prediction, using Shapley attribution to trace the model's reasoning. That kind of interpretability is rare in medical imaging AI, where most synthesis models are black boxes that produce outputs without explaining what drove them. If the synthesis holds up in broader validation, it could extend Alzheimer's staging to clinics that lack PET infrastructure.
The model performed competitively on both reconstruction fidelity and clinically relevant metrics like Braak stage agreement — though "competitive" is not the same as "ready for clinical use," and the validation cohort of 83 subjects is modest by the standards regulators will eventually require.