Pre-Training for Large-Scale Functional Connectome Fingerprinting Supports Generalization and Transfer Learning in...

Functional MRI currently supports a limited application space stemming from modest dataset sizes, large interindividual variability and heterogeneity among scanning protocols. These constraints have made it difficult for fMRI researchers to take full advantage of modern deep-learning tools that have revolutionized other fields such as NLP, speech transcription, and image recognition. To help address these issues, we scaled up functional connectome fingerprinting as a neural network pre-training task, drawing inspiration from speaker recognition research, to learn a generalizable representation of brain function. This approach achieves strong performance for neural fingerprinting on a previously unseen scale, across multiple public fMRI datasets (individual recognition from held out scan sessions: 93% on MPI-Leipzig, 94% on NKI-Rockland, 73% on OASIS-3, and 99% on HCP). Performance is maintained even when evaluation scan duration is truncated to less than two minutes. We show that this representation can also generalize to support accurate neural fingerprinting for completely new datasets and participants of either sex not used in training. Finally, we demonstrate that the representation learned by the network encodes features related to individual variability that partially transfers to new tasks. These results support the development of scalable transfer-learning approaches for future clinical and cognitive neuroimaging applications.Significance statement Deep learning models that leverage the increasing scale of available fMRI data could address fundamental generalization challenges. We drew inspiration from other domains that have successfully used deep learning to address these problems, namely human language technology, to guide our approach for addressing these challenges in neuroimaging. Our pre-training method achieves strong performance for functional connectome fingerprinting, achieving very high recognition accuracy across different tasks, scanning sessions, and acquisition parameters, even when the duration of a scan is limited to less than two minutes. Representations learned by our model could be repurposed to recognize new individuals from new datasets and to predict new participants' cognitive performance and traits.