Learning Generalizable Recurrent Neural Networks from Small Task-fMRI Datasets.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Image Analysis

Background:

• Deep learning excels in image analysis but requires large datasets.
• Medical imaging studies, particularly with task-fMRI, often have limited sample sizes.
• Training effective deep learning models on small datasets is a significant challenge.

Purpose of the Study:

• To develop and evaluate methods for training generalizable recurrent neural networks (RNNs) on small task-fMRI datasets.
• To improve the application of deep learning in medical image analysis for conditions like autism spectrum disorder (ASD).

Main Methods:

• Proposed a resampling method for region-of-interest (ROI)-based fMRI analysis to augment data.
• Incorporated non-imaging variables for subject-specific RNN initialization.
• Selected the most generalizable RNN model using training loss criteria from multiple reinitialized runs.
• Utilized cross-validation to assess model performance on ASD datasets.

Main Results:

• Demonstrated the effectiveness of the proposed methods in training RNNs from small task-fMRI datasets.
• Successfully predicted treatment outcomes in children with ASD (N=21).
• Accurately classified autistic subjects versus typical controls (N=40) using task-fMRI scans.

Conclusions:

• The developed approaches enable effective training of RNNs on limited task-fMRI data.
• These methods facilitate the use of deep learning for analyzing neuroimaging data in conditions with small patient cohorts.
• The study highlights the potential for improved diagnostic and prognostic tools in ASD research using fMRI and deep learning.