AI-PET: Pathology Maps from a Routine MRI
Deep models that synthesize amyloid-β and tau PET from non-invasive T1-weighted MRI, toward low-cost, large-scale Alzheimer's screening.
Overview
In Dr. Jia Guo’s lab at Columbia’s Zuckerman Institute, I lead a project that estimates Alzheimer’s pathology directly from a standard structural MRI. Using deep-learning models, it approximates a picture normally obtained from a costly, less accessible PET scan — from a scan most patients already receive.
Why it matters
PET is the standard way to see the proteins that mark Alzheimer’s, but it needs a radioactive tracer and a scanner few centers have; MRI is comparatively cheap and everywhere. Estimating pathology maps from MRI could make preliminary Alzheimer’s screening far more accessible, and could open up large retrospective studies in cohorts where PET was never acquired.
Related work
Alongside this, I first-authored two abstracts accepted at ISMRM 2026: fast and robust Ktrans mapping of focused-ultrasound blood–brain-barrier opening, and accelerated cerebral blood volume reconstruction from perfusion-weighted MRI.
The project itself is ongoing and not yet published, so I keep the specifics light here.