Assessment of the current and emerging criteria for the histopathological classification of lung neuroendocrine tumors in the lungNENomics project (A17)

Introduction: The lungNENomics project analysed 300 lung neuroendocrine tumors (NETs), of which
259 cases were pathologically reviewed by six expert pathologists.

Aim(s): This dataset assessed the validity of the WHO classification criteria, which diagnose atypical carcinoids based on necrosis and/or two to ten mitoses per 2 mm².

Materials and methods: Expression of two markers of tumor proliferative activity, Ki-67 and phospho-histone H3 (PHH3), was quantified by pathologists and supervised by deep-learning algorithms. In addition, an unsupervised model was trained on WSI H&E to identify potential novel morphological features.

Results: The current classification system suffers from significant inter-observer variability and shortcomings in identifying specifically aggressive cases. A wide range of Ki-67 or PHH3 thresholds can classify lung NETs according to prognosis, with similar sensitivity/specificity ratios to those obtained with the current criteria. Automated scoring of these markers, although faster, leads to comparable conclusions. The unsupervised deep-learning model, although not reproducing the current classification, allows to discriminate cases based on molecular profiles from multi-omics analyses.

Conclusion: Although the mitotic count criterion could be replaced by manual or automated assessment of Ki-67 or PHH3, these markers do not significantly improve the prognostic value or reproducibility of the current classification. Although the potential of morphological features in the current classification of TC/AC seems exhausted, the unsupervised model suggests
features for a more clinically relevant morpho-molecular classification.

Keywords: lung neuroendocrine neoplasm, pathology, deep-learning, ki-67, phh3