No questions here, but you guys might be interested to see the implementation I took with this, where I also used machine learning for Alzheimer's / dementia diagnosis.

https://github.com/jddunn/dementia-progression-analysis

I was using OASIS's public dataset so I only had ~150 images to work with, instead of ~2000. I used transfer learning from ImageNet's dataset to try and get usable results. I also had super limited testing (15-20 patients), but got ~60% accuracy with ~13% false positive rate.

It could be useful to apply those same transfer learning techniques in your team's model.

I find it really interesting that you used machine learning to do the diagnosis given the data you had available - scans seem to be naturally similar to images that we use neural networks on today so t he fit seems good. I was wondering though, did you have any ability to dissect the algorithm after it was trained and see what exact characteristics it was looking for?

The reason I ask is because there's some description of the processes we know indicate Alzheimers, but there could be new signs this algorithm has identified that could be applied more directly?