Medmain Develops AI to Automatically Measure Ki-67 Positive Cell Rate in Immunohistochemical Specimens - Paper Published in Diagnostics -

Medmain Inc. (Headquarters: Fukuoka City, Fukuoka Prefecture; Representative Director and CEO: Osamu Iizuka; hereinafter "Medmain"), provider of the digital pathology support solution "PidPort", has successfully developed an AI that detects Ki-67 positive cells in immunohistochemical digital specimens of the Ki-67 protein and automatically measures the positive cell rate (Labeling Index). In this study, we developed an AI model that automatically distinguishes between Ki-67 positive and negative cells and calculates the positive cell rate by performing cell nucleus detection and cell-level classification on Whole-Slide Images (WSI). The developed AI model demonstrated high classification performance, and a high correlation with evaluation results by pathologists was confirmed. We are pleased to announce that the content of this research has been published in "Diagnostics", issued by MDPI (https://www.mdpi.com). Paper URL: https://www.mdpi.com/3765504 ## Overview of Research Results We have successfully developed an AI that detects Ki-67 positive cells in immunohistochemical digital specimens and automatically measures the positive cell rate (Labeling Index). ## Background of the Research The Ki-67 protein (MKI67), a cell proliferation marker, is expressed in all phases of the cell cycle where cell proliferation occurs (G1, S, G2, and M phases), while it is not expressed in the resting phase (G0 phase) where cell proliferation does not occur. Since Ki-67 shows characteristic expression in the cell nucleus, its expression localization is usually visualized and evaluated using immunohistochemistry (IHC). Ki-67 immunohistochemistry is widely used in pathological diagnosis as an important indicator for evaluating the proliferative capacity of tumors, and the positive cell rate (labeling rate) is evaluated as the Labeling Index (LI). A case with a higher LI indicates faster tumor growth, and it is used in daily clinical practice as a marker for malignancy and prognosis prediction in various tumors such as breast cancer, brain tumors, soft tissue tumors, and neuroendocrine tumors. For example, in breast cancer, it is required to measure the Ki-67 positive cell rate in invasive cancer cells, and it is necessary to count at least 500 to 1000 cells. Conventionally, it is common for pathologists to measure this using a manual counter, but since the measurement requires time and effort, there has been a demand for the development of software that can measure it objectively and efficiently. Against this background, in this study, we developed an AI model that detects Ki-67 positive cells and automatically measures the positive cell rate in Ki-67 immunohistochemical digital specimens. ## Content of the Research We digitized 320 Ki-67 immunohistochemically stained specimens (WSI: Whole-Slide Image) consisting of combinations of various carcinomas, including breast cancer, provided by domestic medical facilities, and created training data by having pathologists annotate Ki-67 positive cells and negative cells. In the developed AI model, after first detecting cell nuclei (Cell detection), positive cells and negative cells are distinguished (Cell classification) by a cell classification model trained using the training data. As a result, we developed an algorithm that automatically counts the number of Ki-67 positive cells and negative cells and calculates the positive cell rate (Labeling Index). In addition, with this AI model, by having a pathologist specify a Region of Interest (ROI) on the Whole-Slide Image, it is possible to detect all the cells included in that region, classify them into Ki-67 positive cells and negative cells, and then automatically measure the positive cell rate. ## Results of the Research In the classification performance of Ki-67 positive cells, a high performance of ROC-AUC 0.981 was obtained. In addition, the Ki-67 positive cell rate (LI) calculated by this AI model showed a high correlation compared with the results of a pathologist evaluating the same image. From these results, it was shown that this AI model is useful as a tool to assist pathologists in evaluating the Ki-67 positive cell rate. ## Future Outlook In this study, we developed AI targeting Ki-67, but the method developed in this study is considered applicable to other proteins (e.g., P53) where positive signals are observed in the nucleus in immunohistochemistry. In the future, we will proceed with the research and development of AI capable of objectively and quantitatively evaluating the expression patterns of not only nuclear proteins but also proteins showing various cellular localizations, such as cell membrane proteins. ## Original Paper - Paper Title: Automated Assessment of Ki-67 Labeling Index Using Cell-Level Detection and Classification in Whole-Slide Images - Japanese Translation: 免疫組織化学デジタル標本におけるKi-67陽性細胞率を自動計測するAIの開発 - Published Journal: Diagnostics - DOI: https://doi.org/10.3390/diagnostics16050816 - Paper URL: https://www.mdpi.com/3765504 ## Authors and Affiliations - Masayuki Tsuneki - Meng Li - Fahdi Kanavati The results of this research were obtained through research and development conducted with the grant from the Fukuoka Prefecture Innovative Medical Device Research and Development Support Project Subsidy. ## Company Overview - Company Name: Medmain Inc. - Establishment Date: January 11, 2018 - Business Description: Planning, development, operation, and sales of medical software and cloud services - Representative Director / CEO: Osamu Iizuka - Locations: - [Tokyo Office] 2F A Aoyama Building, 2-10-11 Minami-Aoyama, Minato-ku, Tokyo - [Fukuoka Office] 104 Chatelet Success, 2-4-5 Akasaka, Chuo-ku, Fukuoka City, Fukuoka - [Silicon Valley Office] 212 Homer Ave, Palo Alto, CA 94301 USA (Medmain USA Inc.) ## Various Related Sites - Corporate Site: https://medmain.com - Product Site: https://service.medmain.com