Comparative analysis of Ki-67 labeling index morphometry using deep learning, conventional image analysis, and manual counting

Abstract

The Ki-67 labeling index is essential for predicting the prognosis of breast cancer and for diagnosing neuroendocrine and gastrointestinal stromal tumors. However, current manual counting and digital image analysis (DIA)-based methods are limited in terms of accurate estimation. This study aimed to assess and compare the capabilities of different DIA systems for Ki-67 counting using the conventional manual counting method. A total of 239 tissue microarray cores from patients with stomach cancer were immunohistochemically stained for Ki-67 and digitally scanned. For the analysis, we employed three different annotation methods: whole TMA core, box selection of the epithelium, and hand-free selection of the epithelium. We used DIA system of 3DHistech, Roche, aetherAI, and manual counting by the pathologists. The annotation methods showed different Ki-67 positivity but were lower than the pathologist manual counting. The results demonstrate that the Roche system is the preferred method for analyzing the entire TMA, whereas aetherAI outperforms the box selection method. Furthermore, 3DHistech is the most accurate method for hands-free selection of the epithelium. The manual counting results showed good agreement among pathologists, with an average intraclass correlation coefficient of 0.93. These results emphasize the importance of carefully selecting annotation methods to determine Ki-67 positivity. To determine the most suitable method for individual laboratories, multiple approaches should be assessed before implementing a DIA system in routine practice.