Transfer learning improves resting-state functional connectivity pattern analysis using convolutional neural...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning models require extensive data for training, which is often unavailable in neuroimaging studies.
• Transfer learning offers a solution by adapting pre-trained deep networks to new tasks with limited data.

Purpose of the Study:

• Investigate the efficacy of transfer learning for age classification and regression using brain functional connectivity.
• Assess the adaptability of deep neural networks trained on large datasets to smaller, distinct neuroimaging datasets.

Main Methods:

• Trained a connectome-convolutional neural network on a large public dataset.
• Applied transfer learning by fine-tuning the pre-trained network on smaller target datasets with varying scanner and protocol parameters.

Main Results:

• Transfer learning enhanced age classification accuracy by approximately 9%-13% on target datasets.
• The approach showed promise in improving the prediction of chronological age, even with limited data.

Conclusions:

• Transfer learning is a viable strategy for adapting deep neural networks to neuroimaging data with sample size limitations.
• This method effectively handles variations in data acquisition and pre-processing protocols.