BLENDS: Augmentation of Functional Magnetic Resonance Images for Machine Learning Using Anatomically Constrained...

Area of Science:

• Neuroimaging
• Deep Learning
• Medical Image Analysis

Background:

• Data augmentation is crucial for deep learning models with limited training data.
• Existing methods for synthesizing four-dimensional (4D) neuroimaging data, like functional magnetic resonance imaging (fMRI), are limited.
• There is a need for validated augmentation techniques to create anatomically realistic 4D fMRI data.

Purpose of the Study:

• To propose and validate a novel data augmentation method for synthesizing realistic 4D fMRI images.
• To enhance the performance of deep learning models in neuroimaging tasks by addressing data scarcity.
• To improve the accuracy of predictive models for clinical applications in neurology.

Main Methods:

• Introduced Brain Library Enrichment through Nonlinear Deformation Synthesis (BLENDS), a new augmentation method.
• BLENDS generates new nonlinear warp fields by spatially blending intersubject coregistration maps.
• Applied generated warp fields to existing 4D fMRI data to create augmented datasets.

Main Results:

• BLENDS successfully generated hundreds of new fMRI images with unique anatomical variations.
• Augmentation significantly improved prediction performance in two neuroimaging tasks.
• For antidepressant response prediction, R-squared increased from 0.055 to 0.103 with 10x augmentation.
• For Parkinson's disease trajectory prediction, R-squared improved from -0.044 to 0.472 with 10x augmentation.

Conclusions:

• The BLENDS method effectively augments fMRI data through nonlinear transformations.
• This augmentation significantly enhances deep learning model performance on clinically relevant predictive tasks.
• BLENDS offers a valuable tool for neuroimaging researchers to overcome data limitations and build more accurate predictive models.