Pulmonary Neuroendocrine Tumor Artificial Intelligence-Assisted Prediction of Postoperative Recurrence from Tumor Biopsy Specimens.

The current WHO (2021) classification of lung neuroendocrine tumors (LNETs) into typical (TC) and atypical carcinoids (AC) relies on resection material and is not applicable to small biopsies. However, accurate risk stratification at the preoperative stage remains clinically important. In this study, we developed an artificial intelligence (AI) model to predict postoperative recurrence in LNET patients directly from hematoxylin and eosin (H&E)-stained tissue fragments, including diagnostic biopsies, independent of the WHO classification. A multiscale vision transformer was trained and evaluated to predict disease recurrence following curative resection, using a population-based cohort of pathologically confirmed stage I-III patients. H&E slides of tissue microarrays (TMAs) from 452 LNET patients (TC = 414, AC = 38; non-recurrence = 381, recurrence = 71) were included. For external validation, an independent TMA cohort of 120 surgically treated LNETs (TC = 91, AC = 29; non-recurrence = 109, recurrence = 11) was utilized. To evaluate clinical applicability, 93 routine preoperative biopsies (TC = 80, AC = 13; non-recurrence = 76, recurrence = 17) were analyzed. The AI model demonstrated robust cross-cohort performance with balanced accuracy of 0.79-0.89 and area under the curve (AUC) of 0.84-0.95. Despite few recurrence cases, the model achieved high sensitivity (0.82-0.94), reliably identifying high-risk patients. Specificity was 0.70-0.85, and negative predictive value (NPV) consistently 0.98 across all datasets, enabling confident recognition of low-risk patients. Overall, the AI model outperformed the WHO-based recurrence prediction, correctly identifying 91/99 (92%) vs. 31/99 (31%) high-risk cases. Performance improvements were statistically significant in the TMA cohorts (Wilcoxon signed-rank test; p < 0.05). This is the first AI-based approach capable of predicting recurrence in LNETs using only H&E-stained biopsies. These findings support AI-driven histopathological assessment as a complementary tool to improve preoperative risk stratification and guide clinical management of LNET patients.