Related Experiment Video
Updated: Aug 5, 2026
05:55
Modeling the Functional Network for Spatial Navigation in the Human Brain
Published on: October 13, 2023
Pre-Training for Large-Scale Functional Connectome Fingerprinting Supports Generalization and Transfer Learning in
Mattson Ogg1, Lindsey Kitchell1
1Research and Exploratory Development Department Johns Hopkins Applied Physics Laboratory.
Eneuro
|July 29, 2026
Summary
Deep learning enhances functional MRI analysis by using functional connectome fingerprinting for scalable, generalizable brain function representation. This method achieves high accuracy even with short scans, paving the way for new neuroimaging applications.
Area of Science:
- Neuroimaging
- Machine Learning
- Computational Neuroscience
Background:
- Functional MRI (fMRI) data is limited by small dataset sizes, high interindividual variability, and diverse scanning protocols.
- These constraints hinder the application of deep learning (DL) in fMRI research, unlike in fields like NLP and image recognition.
- Scalable DL approaches are needed to unlock the potential of large-scale fMRI datasets.
Purpose of the Study:
- To develop a scalable transfer-learning approach for fMRI analysis using deep learning.
- To adapt speaker recognition techniques for functional connectome fingerprinting as a pre-training task.
- To learn a generalizable representation of brain function from fMRI data.
Main Methods:
- Scaled up functional connectome fingerprinting as a neural network pre-training task.
- Applied the approach across multiple public fMRI datasets (MPI-Leipzig, NKI-Rockland, OASIS-3, HCP).
- Evaluated performance on individual recognition, generalization to new datasets/participants, and transferability of learned features.
Main Results:
- Achieved high accuracy in neural fingerprinting across diverse fMRI datasets (e.g., 93-99% on specific datasets).
- Maintained performance even with scan durations under two minutes.
- Demonstrated generalization to completely new datasets and participants, including those of different sexes.
- Showed that learned representations encode individual variability features that partially transfer to new tasks.
Conclusions:
- The developed pre-training method provides a scalable transfer-learning approach for fMRI.
- The learned representations generalize well and can be repurposed for various neuroimaging applications.
- This work supports the advancement of DL in clinical and cognitive neuroimaging, enabling more robust analysis of brain function.
