Thread by @alexisthual

Our paper on using Optimal Transport to compare human cortical surfaces was accepted at #NeurIPS22 🥳
We implement a new OT solver (FUGW) to compare human brains although their anatomy and functional activity patterns differ a lot.
arxiv.org/abs/2206.09398
github.com/alexisthual/fu…

Comparing brains and deriving meaningful statistics from groups of individuals is hard, mostly because (i) brain anatomy varies across individuals and (ii) two individuals presented with the same stimuli will show different activity patterns.

Actually, we even built a web application to get a better intuition of how different brains are across humans.
It allows us to explore the IBC dataset: an in-depth fMRI study of 12 human subjects who have been scanned more than 50 hours each!
Try it now bit.ly/3VpgEcD 🤓

Our idea is simple: we match points across subjects s and t based on functional similarity📊(ie they behave the same way throughout our experiments) while preserving the anatomy of the cortex🧠(two points which are far away in s should be mapped to points that are far away in t)

In order to test derived mappings, we use them to "transport" fMRI maps between a source and a target subject, and observe how correlation between these maps improves compared to other methods (MSM and a baseline method which simply projects individuals onto a template anatomy).

Interesting features of FUGW are that it's fast (~3 minutes to align 2 subjects using 400 fMRI maps at a resolution of 10k points per cortical hemisphere on a V100 GPU) and that it's not too sensitive to changing hyper-parameters🔎

Moreover, while most commonly used alignment methods usually need individual data to lie on the same anatomical template, FUGW can compare individuals on their native anatomies directly. We think this makes FUGW a good candidate for inter-species comparisons 🐵🧑‍🦱

Indeed, it's "unbalanced" feature allows to model shrinkage and/or expansion of cortical areas across pairs of individuals, which we advocate is crucial to model brain differences between species (and even between infants and adults).

But to me, an even more exciting feature is that one can compute the FUGW barycenter of a cohort, ie an "average" individual which minimizes the FUGW loss to all individuals of the group. We observe that our barycenter is more detailed than usual vertex-wise averages.

This work emerged from an exciting and joyful collaboration with the amazing @sixulm1, @nicolas_courty and @RFlamary, my very supportive and masterful advisors @StanDehaene and @BertrandThirion, and the brilliant Tatiana Zemskova!

I would also like to thank @tomamoral, @JallaisMaeliss, Thibault Séjourné and @gabrielpeyre for the useful discussions and coffee breaks, and for their motivational talks and debugging skills😅

Current and future extensions of this work include:
- using features derived from naturalistic stimuli to compare brains💬
- using FUGW draw inter-species comparisons
- scaling🚀! FUGW for resolutions of >100k points per hemisphere (and even volumes?) is on the way

Stay tuned!

We're happy to chat about this online, but if you feel like speaking about this in-person, come and see us
(1) this afternoon at 2:00pm at #NeurIPS22 in Paris in Jussieu or
(2) in New Orleans on Tue 29 Nov 2:30pm, Hall J #915