Alzheimer's: an AI spots signs on brain scans years before diagnosis

Why is Alzheimer's so hard to spot early?

Alzheimer's disease slowly damages the brain and affects memory. Changes begin long before the first visible problems, and diagnosis often comes late.

There is an in-between stage, mild cognitive impairment: small problems with memory or concentration that do not always turn into Alzheimer's. Telling who will get worse and who will not is a real challenge for doctors.

An FDG PET scan is an imaging test that shows how the brain uses sugar, its fuel. Some areas use less as the disease sets in, but these changes are hard to see by eye.

How did the researchers train and test the AI?

The team used 2,109 PET scans from 1,002 patients from ADNI, a large research study that follows people over time. For each patient, the final diagnosis made during follow-up was known: Alzheimer's, mild cognitive impairment, or neither.

They trained a neural network, a program that learns to recognize patterns in images, on 90% of these scans. They tested it on the remaining 10%, then on an independent group of 40 patients scanned between 2006 and 2016.

The AI's results were compared with those of specialist imaging readers looking at the same scans.

What did the study find?

In the independent group of 40 patients, the AI reached an area under the curve of 0.98. This measure runs from 0.5 (chance) to 1 (perfect) and sums up how well a test separates people with the disease from those without.

Set so as not to miss any case, it flagged every patient later diagnosed with Alzheimer's, with 82% specificity, meaning it correctly cleared 82% of people who did not have it. The scans had been taken an average of 75.8 months before the final diagnosis.

The specialist readers spotted four cases out of seven, with 91% specificity. Overall, the AI did better.

What could this mean for patients?

If these results hold up in large groups, an AI could help doctors read brain PET scans and spot earlier the people who need close follow-up.

Knowing early can give time to plan, to join clinical trials or to adapt daily life. But today this tool is not used in routine care, and PET scans are not offered to everyone.

What are the limits of this study?

The independent test group is very small: 40 patients, only seven of whom were diagnosed with Alzheimer's. A near perfect score in so few people may not hold up across thousands.

Most data came from a research study with selected patients, not the general population. The study looks back at past scans rather than using the tool live with new patients. And the AI looked at the whole brain, which makes its reasoning hard to explain to doctors.