Data Augmentation Based on Substituting Regional MRIs Volume Scores.

Area of Science:

• Neuroimaging
• Machine Learning
• Addiction Research

Background:

• Limited training data hinders neural network applications in brain Magnetic Resonance Imaging (MRI) analysis.
• Synthetic data generation is a viable solution to overcome data scarcity in neuroimaging studies.

Purpose of the Study:

• To propose and evaluate a novel data augmentation strategy, regional feature substitution, for neuroimaging analysis.
• To enhance the performance of neural network classifiers in predicting alcohol consumption patterns in adolescents using volumetric MRI data.

Main Methods:

• Developed a data augmentation technique called regional feature substitution.
• Trained a neural network classifier using original and augmented brain MRI datasets.
• Employed 20-fold cross-validation to assess classifier performance on a 5-group classification task.

Main Results:

• Generated over one million synthetic MRI samples from fewer than 500 subjects.
• Achieved 74.1% accuracy in distinguishing baseline alcohol consumption levels (non-drinkers vs. drinkers).
• Obtained a 43.2% weighted accuracy in predicting transitions in alcohol consumption over three years.

Conclusions:

• Regional feature substitution significantly improves classifier performance compared to using original datasets alone.
• The proposed augmentation strategy effectively addresses data limitations in neuroimaging for predicting behavioral transitions.
• This approach holds promise for advancing machine learning applications in psychiatric and neurological research.