The BLENDS Method for Data Augmentation of 4-Dimensional Brain Images

PurposeData augmentation improves the accuracy of deep learning models when training data is scarce by synthesizing additional samples. This work addresses the lack of validated augmentation methods specific for synthesizing anatomically realistic 4D (3D+time) images for neuroimaging, such as fMRI, by proposing a new augmentation method. Materials and MethodsThe proposed method, BLENDS, generates new nonlinear warp fields by combining intersubject coregistration maps, computed using symmetric normalization, through spatial blending. These new warp fields can be applied to existing 4D fMRI to create new augmented images. BLENDS is tested on two neuroimaging problems using de-identified datasets: 1) the prediction of antidepressant response from task-based fMRI in the EMBARC dataset (n = 163), and 2) the prediction of Parkinsons Disease symptom trajectory from baseline resting-state fMRI regional homogeneity in the PPMI dataset (n = 43). ResultsBLENDS readily generates hundreds of new fMRI from existing images, with unique anatomical variations from the source images, that significantly improve prediction performance. For antidepressant response prediction, augmenting each original image once (2x the original training data) significantly increased prediction R2 from 0.055 to 0.098 (p < 1e-6), while at 10x augmentation R2 increased to 0.103. For the prediction of Parkinsons Disease trajectory, 10x augmentation R2 increased from 0.294 to 0.548 (p < 1e-6). ConclusionAugmentation of fMRI through nonlinear transformations with BLENDS significantly improves the performance of deep learning models on clinically relevant predictive tasks. This method will help neuroimaging researchers overcome dataset size limitations and achieve more accurate predictive models.