The bed nucleus of the stria terminalis (BNST) plays a critical role in responses to stress and threat. It has emerged as a key region in anxiety, stress-related disorders, and substance use disorders. However, progress in BNST research has been hindered by the lack of reliable and efficient segmentation methods. To address this gap, we propose nnU-BNST, a fully automated deep learning-based framework leveraging the nnU-Net architecture for accurate BNST segmentation. Our approach entails two novel aspects: (1) it represents the first automated segmentation strategy for the BNST, (2) it systematically investigates the impact of multi-modal MRI inputs on segmentation performance. Three distinct datasets are incorporated for training and evaluation. Experimental results indicate that nnU-BNST consistently outperforms both atlas-based segmentation and other deep-learning-based architectures. Notably, the model achieves its highest accuracy when trained on multi-modal MRI data, yielding a Dice coefficient of 0.768 on a test set containing healthy and trauma-exposed subjects with 19.81% improvement over the best-performing atlas-based approach. Remarkably, although trained exclusively on data from healthy individuals, nnU-BNST exhibits strong generalization to people with trauma-exposure. These findings establish nnU-BNST as a robust and effective tool for subject-specific BNST segmentation, with significant potential to advance research on clinically relevant brain regions.