Pancreatic cancer is an aggressive malignancy with an extremely poor prognosis, where lymph node metastasis (LNM), as a critical prognostic factor, is closely associated with patient recurrence and overall survival. In clinical practice, preoperative assessment of lymph node status is critical for treatment decisions. However, diverse acquisition conditions, scarce data resources, and limited utilization of 3D information pose challenges for imaging-based preoperative assessment algorithms. In this work, we propose a generalizable automated method for LNM prediction in pancreatic cancer. Specifically, to address the variability in sampling conditions, a sampling simulation generation model is proposed. This model simulates the acquisition process of enhanced CT to generate latent modalities, thereby improving the model adaptability to diverse data. Then, a spiral transformation is introduced to project the 3D CT images onto a 2D plane, which fully leverages the 3D information and enhances the representation of tumor features. Finally, class-guided contrastive learning is developed, which clusters together the features of each class, thereby enhancing the discriminative of the learned fine-grained features. Extensive experiments demonstrated that this method achieved remarkable prediction performance, while its stability and generalizability were validated through independent tests. Therefore, this method provides a valuable potential clinical tool for predicting LNM in pancreatic cancer.