Assessment of the current and emerging criteria for the histopathological classification of lung neuroendocrine tumours in the lungNENomics project

Abstract

Background: Six thoracic pathologists reviewed 259 lung neuroendocrine tumours (LNETs) from the lungNENomics project, with 171 of them having associated survival data. This cohort presents a unique opportunity to assess the strengths and limitations of current World Health Organization (WHO) classification criteria and to evaluate the utility of emerging markers.

Patients and methods: Patients were diagnosed based on the 2021 WHO criteria, with atypical carcinoids (ACs) defined by the presence of focal necrosis and/or 2-10 mitoses per 2 mm². We investigated two markers of tumour proliferation: the Ki-67 index and phospho-histone H3 (PHH3) protein expression, quantified by pathologists and automatically via deep learning. Additionally, an unsupervised deep learning algorithm was trained to uncover previously unnoticed morphological features with diagnostic value.

Results: The accuracy in distinguishing typical from ACs is hampered by interobserver variability in mitotic counting and the limitations of morphological criteria in identifying aggressive cases. Our study reveals that different Ki-67 cut-offs can categorise LNETs similarly to current WHO criteria. Counting mitoses in PHH3+ areas does not improve diagnosis, while providing a similar prognostic value to the current criteria. With the advantage of being time efficient, automated assessment of these markers leads to similar conclusions. Lastly, state-of-the-art deep learning modelling does not uncover undisclosed morphological features with diagnostic value.

Conclusions: This study suggests that the mitotic criteria can be complemented by manual or automated assessment of Ki-67 or PHH3 protein expression, but these markers do not significantly improve the prognostic value of the current classification, as the AC group remains highly unspecific for aggressive cases. Therefore, we may have exhausted the potential of morphological features in classifying and prognosticating LNETs. Our study suggests that it might be time to shift the research focus towards investigating molecular markers that could contribute to a more clinically relevant morpho-molecular classification.

Keywords: Ki-67; PHH3; deep learning; histological classification; lung neuroendocrine tumours.

Figure 1

Presentation of the lungNENomics cohort and protocols for pathological review and digital pathology analyses. (A) Study design (created with BioRender.com). (B) Flow diagram of the central pathological review leading to final diagnoses. (C) Kaplan–Meier curves of RFS according to the reference diagnoses. Grey dashed lines indicate median survival; P value corresponds to the log-rank test. AC, atypical carcinoid; HE, haematoxylin and eosin; PHH3, phospho-histone H3; RFS, recurrence-free survival; TC, typical carcinoid; WSI, whole-slide image.

Figure 2

Assessment of number of mitoses counted on HE/HES slides for the classification of LNETs. (A) Reference WHO LNET classification system. (B) Distributions of the number of mitoses per 2 mm² by reader based on reference diagnosis. (C) Prognostic value of TC and AC groups for different mitotic count thresholds. Left panel: forest plot for AC versus TC groups based on RFS. Elements are coloured if the Wald test is significant (P < 0.05). The error bars around the hazard ratios correspond to the 95% CI. When the error bars end in an arrow, this means that the confidence intervals are associated with a large value. For ease of reading, this value is not shown and the arrow indicates that the true value is to the right. The name of the group in bold in the panel titles is the target group for which hazard ratios are reported in comparison to the reference group, which is not written in bold. Middle panel: forest plot for AC versus reference AC based on PFS. Right panel: percentages of cases diagnosed as TC and AC. AC cases diagnosed as such because of the presence of necrotic foci are labelled as AC.Necrosis. The proportion of NA represents cases with no majority vote for TC or AC (three votes each). AC, atypical carcinoid; CI, confidence interval; HE, haematoxylin and eosin; HES, haematoxylin, eosin and saffron; LNET, lung neuroendocrine tumours; NA, not available; Nb, number; PFS, progression-free survival; RFS, recurrence-free survival; TC, typical carcinoid; WHO, World Health Organization.

Figure 3

Effect of including Ki-67 or PHH3 assessment in the LNET classification system. First column: hypothetical classification system. Individual pathologists’ diagnoses were combined by majority vote, meaning that some cases may remain unclassified if observations resulted in three AC and three TC diagnoses. The four hypothetical classification systems are the following: (A) Based on manual assessment of Ki-67 expression. (B) Based on automatic assessment of Ki-67 expression. (C) Based on manual mitotic count on regions with PHH3 expression. (D) Based on automatic assessment of PHH3 expression. Second to fourth columns: effects of the hypothetical classification system on the prognostic value of the two resulting groups. The significance of the Wald tests associated with the Cox models described on the left is colour-coded for the different thresholds of the variable studied. Fifth column: percentages of cases diagnosed as TC or AC according to the hypothetical classification system. AC cases with foci of necrosis diagnosed by more than three pathologists were labelled AC.Necrosis. The name of the group in bold in the panel titles is the target group for which hazard ratios are reported in comparison to the reference group, which is not written in bold. (E) Left panel: For each classification system, the coloured bar corresponds to the 10-year RFS rate based on the diagnoses resulting from majority voting; these rates are associated with the 95% confidence intervals. Middle panel: percentages of cases diagnosed as TC or AC according to the hypothetical classification system resulting from majority voting. Right panel: Harrell’s C-index, also known as the correspondence index, comparing the quality of univariate Cox models for RFS incorporating the different diagnoses resulting from the classification system mentioned on the left. Error bars around the estimator correspond to confidence intervals. Vertical lines indicate the boundary of the confidence interval resulting from the reference classification system used for comparison. Coloured dots indicate the results obtained from Cox models built on the diagnoses resulting from each pathologist’s observations (one model for each pathologist). AC, atypical carcinoid; HR, hazard ratio; LNET, lung neuroendocrine tumours; PFS, progression-free survival; PHH3, phospho-histone H3; RFS, recurrence-free survival; TC, typical carcinoid; NA, not available.

Figure 4

Unsupervised deep learning experiment on HE/HES WSI of LNET patients. (A) Two-dimensional morphological map resulting from the uniform manifold approximation and projection (UMAP) dimensionality reduction technique applied to Barlow Twins-encoded vectors, each tile is coloured according to the reference diagnosis of the patients. (B) Representation of the 18 communities on (A). Some communities are annotated according to enrichment for certain morphological features, in line with the annotations of pathologists on the WSI. (C) Proportion of tiles in each Leiden community by tumour type. The horizontal black line represents the total proportion of TC versus AC tiles included in the experiment. At the top of the bar, the presence of a star indicates whether a community is significantly enriched for a type; the colour of the star indicates for which tumour types it is enriched. (D) Kaplan–Meier curves of PFS as a function of predicted diagnoses using random forest. The purple curves correspond to the diagnoses with the highest probability between the two types. The green curves correspond to the diagnoses predicted if the 50 most likely ACs were classified as such to obtain a group of the same size as the reference diagnosis. The P values of the log-rank test associated with the type of predictions are shown in the legend. AC, atypical carcinoid; HE, haematoxylin and eosin; HES, haematoxylin, eosin and saffron; LNET, lung neuroendocrine tumours; STAS, spread through alveolar spaces; TC, typical carcinoid; WSI, whole-slide image.