Combining Deep Learning Networks with Permutation Tests to Predict Traumatic Brain Injury Outcome

Reliable prediction of traumatic brain injury (TBI) outcome using neuroimaging is clinically important, yet, computationally challenging. To tackle this problem, we developed an injury prediction or classification pipeline based on diffusion tensor imaging (DTI) by combining a novel deep learning approach with statistical permutation tests. We first applied a multi-modal deep learning network to individually train a classification model for each DTI measure. Individual results were then combined to allow iterative refinement of the classification via Tract-Based Spatial Statistics (TBSS) permutation tests, where voxel sum of skeletonized significance values served as a classification performance feedback. Our technique combined a high-performance machine learning algorithm with a conventional statistical tool, which provided a flexible and intuitive approach to predict TBI outcome.