Unveiling Adolescent BNST Development with Deep Learning: Implications for Alcohol Use and Negative Affect

About This Grant

Abstract The bed nucleus of the stria terminalis (BNST) plays a pivotal role in processing emotional information and is linked to key brain regions relevant to mental health and alcohol use disorders (AUD). Despite decades of foundational studies emphasizing BNST's significance in anxiety, stress, and addiction, these studies have primarily been cross-sectional, often relying on small datasets comprising only a few dozen individuals. Both negative affect (anxiety and depression) and problematic drinking typically emerge during adolescence, highlighting the importance of investigating the BNST during this stage of development. Understanding the development of BNST during adolescence has been hampered by significant time required to manually trace the BNST. To address this barrier, we will develop automatic BNST segmentation algorithm to identify BNST and acquire BNST measures in the NCANDA dataset, facilitating investigation into BNST development in adolescents and the impact of BNST changes on the emergence of anxiety and alcohol consumption. Our central hypothesis is that developmental increases in BNST volume and function will be associated with increased alcohol use and heightened negative affect during adolescence. In Aim 1, we will develop the first automatic BNST segmentation model via deep learning. Our work will address two key challenges in BNST segmentation through deep learning: the scarcity of annotated samples and the inconsistency in segmented BNSTs from diverse sources. In Aim 1-A, we plan to generate synthetic brain-segmented BNST pairs using a generative adversarial network and transfer learning for data augmentation. This method aims to closely replicate the distribution of real data, facilitating the creation of a more comprehensive training dataset for subsequent deep learning-based segmentation. In Aim 1-B, we will construct an automatic BNST segmentation model with label fusion through deep learning. The label fusion net in this model will integrate individually segmented BNSTs obtained through deep learning with multiple BNST atlases, enhancing the overall accuracy of segmentation. In Aim 2, we will investigate whether the increase of BNST volume and function can predict increases in alcohol drinking or negative affect during adolescence. Aim 2-A will examine whether the increase BNST volume and function at initial visit can predict increases in alcohol use or negative affect. Aim 2-B will measure BNST volume and function developmental trajectories and determine whether they are positively correlated with increases in alcohol use or negative affect. The study's impact lies in filling critical knowledge gaps regarding BNST development in adolescents and the impacts BNST development on alcohol use and negative affect. Insights gained could inform targeted interventions to mitigate alcohol-related risks during adolescence.