Active learning seeks to quickly improve model performance while reducing expert annotation effort, which is especially valuable in medical imaging.We study whether an uncertainty-based active learning scheme scales to a large, heterogeneous data collection and what performance is achievable with a limited annotation budget for nipple segmentation in breast magnetic resonance imaging – a precursor task to report lesion localization that has seen little automation so far.We prospectively evaluated an iterative active learning pipeline on 3,762 examinations from nine institutions / datasets, combining a 3D U-Net with Monte Carlo dropout-based image-wise uncertainty and a simple diversity-aware selection strategy. For almost all institutions, the nipple center-of-gravity error and Dice score improve substantially during the first three iterations, before largely plateauing, indicating diminishing returns. Uncertainty decreases in tandem with performance improvements, making it a practical proxy to guide annotation stopping. Persistent challenges involve absent or ambiguous nipples and institution-specific artifacts. Our results demonstrate that uncertainty-driven active learning can efficiently improve nipple segmentation on multi-center MRI with limited annotations. They highlight the need to detect nipple presence before segmentation and show that even if a model improves on data from most institutions when adding more data to the training, it can still achieve worse results in other institutions.