Abstract <p>Deep convection in the Labrador Sea is a key process sustaining the Atlantic Meridional Overturning Circulation (AMOC), and its variability is strongly influenced by mesoscale eddies, in particular Irminger Rings (IRs). Accurate detection of IRs is therefore essential for quantifying their role in modulating convection. In this study, we present and evaluate two approaches for IR detection in high-resolution ocean simulations of the Subpolar North Atlantic. First, we enhance a traditional local-extrema algorithm by applying Bayesian optimization to objectively tune its parameters, improving detection quality and reducing subjectivity. Second, we develop a convolutional neural network, inspired by U-Net, which is pretrained on heuristic detections and subsequently fine-tuned with expert-labeled data. This two-stage training strategy substantially increases detection skill, with the fine-tuned network clearly outperforming both the heuristic baseline and its optimized variant. The demonstrated accuracy and robustness of the U-Net approach make it a promising foundation for constructing trajectory databases of IRs in future work, enabling more reliable studies of eddy–convection interactions and their implications for AMOC variability.</p>