Leveraging recent advances in explainable AI to decode cognitive functioning from T1/FLAIR weighted MR images in MS - An international federated learning approach

An important and long-standing problem in multiple sclerosis (MS) is the clinico-radiological paradox:
the lack of a strong link between brain damage seen through MR imaging and clinical parameters
that capture physical or cognitive impairment.
We hypothesize that recent advances in artificial intelligence (AI) will enable us to reduce this
paradox by elucidating pathophysiological measurements that are not captured through standard
brain volumetry.
More specifically, we will use deep learning (DL) to decode cognitive impairment in MS. However,
deep learning requires large datasets to train the model. For reasons of data protection, such large
datasets are not available. We have solved this by setting up a Federated Learning network, already
including 1400 subjects with MS - that we will further extend to at least 2700 within this project.
Further, the translation to the clinic is limited as these models are perceived as black boxes by
clinicians. Therefore, I will build upon recent advances in explainable AI (XAI) to unravel the decision
process. The gained insights will form the corner stone of a Delphi study in which I will investigate
what is needed to trust AI models.
In summary, this project will combine DL methods to [1] solve the cognitive paradox in MS by
identifying new damage patterns in MRI, [2] unravel how DL models can reach the clinic and [3]
deliver an extended international federated learning network that can deliver trusted DL models in
the MS field.