This study looks at how deep learning could be used to predict internalizing problems from brain structure in kids. The magnetic resonance images (MRIs) that were used in this study were collected from four independent datasets: the Adolescent Brain Cognitive Development (ABCD) study, the Healthy Connectome Project Development (HCP-D) study, and the Province of Ontario Neurodevelopmental (POND) network. A binary classification model was built that used brain structures as input to determine whether they were clinically or non-significant to internalizing problems. Of the 14,523 children, 14% were found to have clinically significant internalizing problems, and these were more prevalent in kids with a neurodevelopmental condition (31%) compared to those without one (8%). For the longitudinal analysis, 19% of patients had worsening changes in internalizing problems between the baseline and follow-up. Like the cross-sectional analysis, patients with an ND diagnosis were more likely to have a worsening trajectory (17%) compared to those without one (20%). For cross-sectional analysis, the brain structures that had the most impact on the prediction were the right frontal lobe, areas of the left temporal lobe, left rostral middle frontal, right medial orbitofrontal gyri, and posterior corpus callosum. For the longitudinal model the brain structures that had the most impact were the areas of the right temporal lobe, left and right lingual gyrus, and the brainstem. In the discussion, one of the main points discussed was how the deep learning model did a better job cross-sectionally achieving an AUC of 0.80, while longitudinally the model preformed sub-optimally achieving an AUC of 0.60. Another point that was hypothesized is that biomarkers associated with ND more reliably predict mental health difficulties.