Whenever I see results like this, I wonder if these scientists can diagnose depression with an MRI. It would be pretty easy to do that experiment, and my sense is that we just aren't there yet. And so I wonder if experiments like this are just measuring noise.
The cluster analysis they used means that we could train a CNN or transformer model to classify and it would be straightforward. I mean you can classify based on the clusters already.
A transformer model would be interesting because you can map voxels to tokens and then look for patterns there. The embeddings alone would be interesting.
Then there are generative AI approaches that can be integrated in as well.
There's a few free ideas for anyone interested. Neat stuff!
