An FDG-PET scan is a spending report. Inject a sugar analogue, wait, and the camera records where the brain is burning fuel. In a person living with Alzheimer's disease, parts of that report dim in a recognisable order, and a trained reader can see it. For the past decade machine learning has been able to see it too, often with impressive accuracy, and almost always without being able to say what it looked at. That silence is the whole problem. A clinician cannot act on a number whose reasoning is sealed, and a family cannot be told "the network decided" and feel they have been told anything.
The paper, from C. Vazquez-Garcia, F. J. Martinez-Murcia and colleagues at the University of Granada with co-authors in Valencia and Malaga, appeared in NeuroImage (online 12 March 2026, issue of 15 April 2026) using data from the Alzheimer's Disease Neuroimaging Initiative. It is a modelling study on existing scans, not a trial and not a prospective test in patients. They built a variational autoencoder, a model that squeezes a whole brain volume down to a handful of numbers and then tries to rebuild the image from them, and they compressed to just eight. The twist is semi-supervision: during training they pushed the first of those eight numbers, which they call z0, to correlate with ADAS13, a standard cognitive assessment score, and the second to correlate with age. The dataset was 3,466 FDG-PET scans from 1,632 people, divided by participant into 1,060 for training, 244 for validation and 328 held back for testing, with mean ages between 72.8 and 73.8 years.
On the held-out test scans, z0 lined people up along a continuum of severity: the authors report an absolute correlation of 0.790 with ADAS13 and 0.810 with average FDG-PET uptake, both at p much less than 0.001. Rerunning the whole thing with Spearman rank correlation instead of Pearson inside the regularisation term gave 0.77, which is the kind of check that suggests the result is not an artefact of one statistical choice. Then the biology test: z0 was compared against structural measures nobody had told it about. It tracked hippocampal volume at 0.48, medial temporal lobe volume at 0.45, entorhinal cortex at 0.37 and fusiform gyrus at 0.34, all volumes normalised for brain size. The gradient matches the known anatomical order of Alzheimer's progression, strongest where the disease starts.
The explainability comes next, and it is the part worth dwelling on. Because the model is generative, the authors could dial z0 up and down and decode average brain images at each level of cognitive impairment, then run a voxel-wise general linear model over those synthetic brains to ask, plainly, where did the fuel go. The answer: reduced metabolism predominantly in the hippocampus, and across major resting state networks, particularly the default mode network and the central executive network. None of that is new biology. That is the point. The model's reasoning was rendered region by region and checked against what three decades of PET literature already says, which is how you earn the right to trust it on something it has not seen before.
The quieter achievement is what the other six latent dimensions ended up holding. They absorbed affine transformations, rotation, translation and scaling, along with intensity variation: head position in the scanner, differences between imaging sites, the ordinary noise of multi-centre data. The authors note the effect is visible in things as mundane as the orientation of the tentorium cerebelli, the membrane separating cerebrum from cerebellum. In other words, the framework separated the disease signal from the furniture of the scan. For anyone who has wondered whether an imaging algorithm is reading pathology or reading the make of the machine, that separation is the grammar underneath the claim.
For a family, the honest translation is narrow. This is not a new scan to ask for, not a score anyone can be given, and not evidence that FDG-PET is better or earlier at anything than it was last year. ADNI is a research cohort recruited for Alzheimer's studies, not a cross-section of everyone who walks into a memory clinic, and the paper reports correlations and region maps rather than sensitivity, specificity or accuracy against a clinical diagnosis. What it offers is infrastructure: a way of building imaging models that can be interrogated afterwards, which is the condition regulators and clinicians will reasonably demand before any such tool is allowed near a consultation room.