A multi-center study using a deep learning Mamba model analyzing pancreas and liver CT scans accurately predicted early postoperative liver metastases in pancreatic cancer and identified patients likely to benefit from neoadjuvant therapy. The work also highlights how imaging-based models may outperform conventional risk factors for early recurrence prediction. If externally validated, the approach could support more precise treatment selection—particularly for patients where early metastatic risk is difficult to quantify pre-operatively. For oncology development, it adds momentum to model-driven stratification beyond pathology and clinical staging alone.