Understanding Human–AI Discrepancy in Breast Cancer TIL Assessment: A Multi-Rater and Perceptual Bias Study

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Objective: Tumor-infiltrating lymphocytes (TILs) in breast cancer are one of the most important indicators of the immune response within the tumor microenvironment. They play a particularly significant prognostic and predictive role in triple-negative and HER2-positive subtypes. However, substantial inter-observer variability has been reported in TIL scoring among pathologists, which limits its reliability in clinical practice. The aim of this study was to evaluate the agreement between artificial intelligence (AI) models and pathologists in TIL scoring and to compare this agreement using different statistical approaches, thereby assessing the potential of AI integration into pathology practice. Materials and Methods: Digitized histopathological images of breast cancer cases were included in the study. Tumor regions annotated by pathologists were evaluated for both stromal TIL percentage and the proportion of stromal tumor area within each ROI, with assessments performed independently by three pathologists and two AI models. Agreement was assessed among pathologists, between pathologists and AI, and between AI models. Statistical analyses included intraclass correlation coefficient (ICC), Cohen’s and Fleiss’ kappa, correlation tests, and Bland–Altman analysis. In addition, categorical agreement was examined using different cut-off values. Results: Inter-pathologist agreement was high, with an ICC of 0.81. In contrast, the global agreement between pathologists and AI models was lower (ICC 0.41). Pairwise comparisons of pathologist–AI agreement yielded substantially lower ICC values (0.12–0.21), although this improved to 0.53 when three pathologists were assessed jointly with a single AI model. The strongest categorical agreement was observed with dichotomized TIL scores (≤10% vs. >10%), whereas multi-category classifications were associated with a marked reduction in kappa values. Spearman correlation coefficients between pathologists and AI models ranged from moderate to good (ρ = 0.48–0.81). Agreement between the two AI models themselves was moderate, with an ICC of 0.64. Conclusion: This study demonstrated that AI models were able to show a certain degree of correlation with pathologists in TIL scoring but remained insufficient in terms of absolute agreement. Differences were more pronounced at higher TIL scores. These findings highlight the need for standardized algorithms trained on larger and more diverse datasets before AI-based TIL scoring can be reliably implemented in clinical practice.

This publication has 39 references indexed in Scilit: