Improving the accuracy of reporting Ki-67 IHC by using an AI tool

Abstract

Ki-67 proliferative index (PI) scoring is measured by estimating the proportion of the number of active cell nuclei in hotspot regions within immunohistochemical (IHC) stained slides. It provides valuable information about the rate of proliferation in a tumour. Manual scoring of Ki-67 PI is laborious, time-consuming and often the victim of interobserver variability between pathologists. This motivated us to develop an AI-based method to automate Ki-67 PI scoring with the aim to improve the concordance of pathologists' inter-observability through aided diagnosis. We sourced 88 sequential cases of sarcomas for our study. We applied watershed algorithm to perform nuclear segmentation on 440 regions of interest (ROI). A study was conducted where three pathologists scored the Ki-67 PI on the ROIs with and without AI-assistance. Our study demonstrated great concordance between the pathologists scoring with AI-assistance. After AI assistance, inter-pathologist discordance was significantly reduced by 82.1 %

Keywords

• Artificial intelligence
• Machine learning
• Ki-67
• Sarcoma
• Inter-observer variability
• Concordance