A new, machine learning-based approach to metastatic neuroendocrine tumors of unknown origin.

Neuroendocrine tumors (NETs) frequently present at a metastatic stage, particularly with liver metastases. Identifying the site of the primary tumor is critical for guiding therapy but often proves difficult. Small intestine NETs are especially distinct in their prognosis and treatment. To address this challenge, we developed a novel, machine learning-based tool to predict the site of origin-specifically small intestine or pancreas-using routine hematoxylin and eosin (H&E)-stained slides from hepatic metastases. To avoid mislabeling in the clinically relevant scenario of any possible tumor origin, the method applies a two-step approach with optional abstention for uncertain classifications or non-small intestine/non-pancreas cases. In a retrospective, clinically realistic cohort with unrestricted tumor origin, the model identified small intestine NETs with a sensitivity of 71.4% at 100% specificity and positive predictive value (PPV), and high negative predictive value. A relevant subset of pancreatic NETs can also be reliably detected (sensitivity 33.3%, specificity 94.1%, PPV 85.7%). Generalizability and robustness were rigorously validated on an external dataset using different scanners, institutions, and resection techniques. The tool is intended as an additional method where other diagnostic modalities remain inconclusive regarding the location of the primary tumor. To facilitate further research and clinical translation, all models and extracted features are publicly released.