Background/Objectives: Although Ki-67 is not formally incorporated into the grading system of pulmonary neuroendocrine neoplasms (PNENs), it is widely used as an adjunct marker to reflect proliferative activity and support diagnostic stratification. Manual Ki-67 assessment is subject to interobserver variability and methodological limitations. This study aimed to evaluate the reliability and performance of two artificial intelligence (AI)-based image analysis systems in Ki-67 index assessment and to compare their results with expert pathologist evaluation in pulmonary neuroendocrine tumors. Methods: A total of 63 pulmonary neuroendocrine neoplasm cases, including typical carcinoid (n = 29), atypical carcinoid (n = 13), and large cell neuroendocrine carcinoma (n = 21), were retrospectively analyzed. Ki-67 proliferation indices were independently assessed by four pathologists within predefined hotspot regions, counting approximately 2000 tumor cells per case. The same regions were analyzed using two AI-based image analysis systems (Roche uPath Ki-67 and Virasoft Virasight Ki-67). Interobserver agreement among pathologists was evaluated using the intraclass correlation coefficient (ICC), and concordance between manual and AI-based assessments was assessed using Spearman’s correlation and linear regression analyses. To account for potential scanner/platform effects, slides were digitized using two different whole-slide scanners (VENTANA DP® 600 and Leica Aperio AT2), and color normalization and quality control procedures were applied prior to AI-based analysis. For clinical interpretability, Ki-67 indices were stratified into categorical groups based on tumor subtype-specific thresholds (0–25%: high), and agreement between manual and AI-based categorical scoring was evaluated using Cohen’s kappa coefficient. Results: Among the 63 pulmonary neuroendocrine neoplasm cases, Ki-67 proliferation indices varied across tumor subtypes, with typical carcinoids showing low, atypical carcinoids intermediate, and large cell neuroendocrine carcinomas high proliferative activity. Interobserver agreement among four pathologists was excellent (ICC = 0.998, 95% CI: 0.996–0.998). Strong correlations were observed between manual Ki-67 assessments and AI-derived indices, with Spearman correlation coefficients of 0.961 (95% CI: 0.918–0.982) for Roche AI and 0.904 (95% CI: 0.821–0.949) for Virasoft AI, and 0.926 (95% CI: 0.842–0.968) between the two AI systems. Bland–Altman analyses demonstrated minimal mean differences and most cases within the 95% limits of agreement, indicating high concordance without systematic bias. Categorical agreement analysis, using subtype-specific Ki-67 thresholds (0–25%: high), showed excellent concordance between manual and AI-based scoring (Cohen’s kappa 0.877 for Roche AI and 0.827 for Virasoft AI; p < 0.001), confirming the clinical interpretability and reproducibility of AI-based Ki-67 assessment. Conclusions: AI-based Ki-67 index assessment shows strong concordance with expert pathologist evaluation and reflects biologically relevant differences among pulmonary neuroendocrine neoplasm subtypes. These results suggest that AI-assisted Ki-67 analysis may serve as a reproducible and objective adjunct to routine diagnostic practice in pulmonary neuroendocrine tumors.
Teoman et al. (Wed,) studied this question.