Peak structures in the scattering experiments are often interpreted as a manifestation of a resonance state. A more rigorous approach requires that we associate at least one S-matrix pole to the peak structure to qualify as a physical state. The presence of a nearby threshold complicates the situation. Sometimes, near-threshold virtual and quasi-bound states produce a similar peak structure. In our work, we phrased the problem as a classification task and solved them using deep learning. In the first part of this talk, I will discuss how deep learning is applied to distinguish virtual and bound state enhancements in the single-channel nucleon-nucleon system. Even without appealing to deuteron’s existence, our deep neural network models can distinguish the two enhancements. In the last part, I will discuss how this approach can be extended to the coupled-channel problem. Specifically, we designed a deep neural network that can extract the coupled-channel S-matrix pole configuration. The resulting training dataset requires a nonconventional training loop; otherwise, the deep neural network will not learn. Finally, we apply our model to the study of the pion-nucleon scattering near the eta-nucleon threshold. I will also discuss some preliminary results.