Automated Assessment of Ki-67 Labeling Index Using Cell-Level Detection and Classification in Whole-Slide Images

Background: The Ki-67 labeling index (LI) is a widely used marker of tumour proliferation, yet its manual assessment is time-consuming and subject to substantial inter-observer variability. Automated methods may improve reproducibility, but their clinical relevance depends on achieving performance comparable to expert pathologists. Method: We evaluated an artificial intelligence (AI)-based, cell-level system for automated Ki-67 LI assessment that detects and classifies individual tumour cell nuclei as Ki-67-positive or -negative. After nuclear detection using a pre-existing cell detection model, a lightweight convolutional neural network classifier operating on nucleus-centred patches was trained, and then applied to cases independently assessed by three pathologists. Agreement between AI-derived and human Ki-67 LI values was compared directly with inter-pathologist agreement across a range of proliferation levels. Results: The AI-based cell classification achieved 98% AUC on a test set consisting of 71K positive and 170K negative image patches centred on nuclei. On the automated Ki-67 LI assessment, the AI system demonstrated concordance with expert pathologists comparable to human inter-observer variability. Conclusions: These results support the potential of cell-level automated Ki-67 assessment as a reproducible decision-support tool for routine histopathological practice.