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

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 advantage of modern deep-learning tools that have revolutionized other fields such as NLP, speech transcription, and image recognition. To 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 sets a new high-water mark for neural fingerprinting on a previously unseen scale, across many popular public fMRI datasets (individual recognition over held out scan sessions: 94% on MPI-Leipzig, 94% on NKI-Rockland, 73% on OASIS-3, and 99% on HCP). Near-ceiling performance is maintained even when the duration of the evaluation scan 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 not used in training. Finally, we demonstrate that the representation learned by the network encodes features related to individual variability that supports some transfer learning to new tasks. These results open the door for a new generation of clinical applications based on functional imaging data. SIGNIFICANCE STATEMENTDeep 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 AI to address these problems, namely human language technology, to guide our exploration of the potential for this approach in neuroimaging. Our pre-training approach sets a new high-watermark 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. We showed that we could re-purpose the representation learned by our model to recognize new individuals from new datasets and to predict new participants cognitive performance and traits.