Multimodal Deep Learning for Pediatric Mild Traumatic Brain Injury Detection

Despite its prevalence, little is known about the pathophysiology of mild traumatic brain injury (mTBI). This makes it difficult for clinicians to accurately diagnose mTBI and to predict outcomes in affected children, thereby highlighting the urgent need to identify novel and efficacious biomarkers of pediatric mTBI. To address this important knowledge gap, this study introduced a multimodal magnetic resonance imaging (MRI) based deep learning approach toward the classification of mTBI as compared with mild orthopedic injury (OI) by considering both structural MRI (sMRI) and diffusion tensor imaging (DTI). Firstly, convolutional features were extracted by employing a pre-trained DenseNet to capture the morphological features of both modalities. Next, by employing Deep Canonical Correlation Analysis (DCCA), distinct features obtained from the sMRI and DTI data were integrated into a multi-modal embedding. The obtained DCCA fused compact multimodal features were then fed to a random forest (RF) classifier that was used to classify mTBI versus mild OI. Additionally, to visualize the intra-individually heterogeneous brain regions that DenseNet most heavily relied upon for making classification, Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to the DenseNet outcomes for both modalities. According to the experimental outcomes on the clinical dataset, the introduced multimodal deep learning strategy improved the classification accuracy by 8.6% (from 75.8% to 84.4%) and 7.8% (from 76.6% to 84.4%) when compared to the unimodal morphological features, as generated from sMRI and DTI.