Lunit "AI Enables Universal Analysis of Immunohistochemical Staining"… Published in Nature Sister Journal

Medical AI company Lunit announced on the 12th that research proving the excellent performance of its AI-based immunohistochemistry (IHC) analysis solution, 'Lunit SCOPE uIHC,' has been published in 'npj Precision Oncology (IF 6.8[1]),' a sister journal of the world-renowned scientific journal Nature.

Immunohistochemistry (IHC) is a method used to detect specific proteins or antigens within cancer tissues. With the recent surge in new anticancer drug development, the importance of IHC, which evaluates the expression levels of anticancer drug targets, has increased. Previously, pathologists manually assessed these, but there were limitations due to inter-observer variability and subjective interpretation.

Deep learning-based AI analysis has emerged as an alternative; however, developing AI models requires large-scale datasets, and especially for new immunohistochemistry types used in novel drug development, obtaining large datasets has been challenging, limiting AI model development.

Lunit’s research team collaborated with multiple large hospital research teams in Korea to overcome these limitations by validating the performance of a universal immunohistochemistry (UIHC) analysis AI model.

The research team utilized slide images from a total of 3,046 patients, including PD-L1 protein staining data from lung, bladder, and breast cancer tissues, as well as HER2 protein staining data from breast cancer. They compared the performance of single-cohort AI models developed with each dataset and a multi-cohort AI model trained on integrated datasets. The multi-cohort AI model trained on all four datasets (Lunit SCOPE uIHC) demonstrated the best performance.

Specifically, the agreement index (Cohen's Kappa) between pathologist readings and AI analysis results was 0.792 for trained cancer types and immunohistochemistry types, surpassing the highest agreement of 0.744 from single-cohort AI models. For analysis of untrained new cancer types and immunohistochemistry types, it recorded an agreement of 0.610, exceeding the 0.508 of models trained on single datasets by over 10%.

In the evaluation of the Tumor Proportion Score (TPS), which indicates the proportion of tumor cells positive for immunohistochemical staining among all tumor cells, Lunit SCOPE uIHC achieved 75.7% accuracy and an AUC[2] of 0.921, demonstrating superior analytical capability compared to other AI models.

The AI model trained on diverse data is significant in proving its capability to analyze multiple cancer types and immunohistochemistry stains. It can analyze new cancer types or immunohistochemistry types without additional training, reaffirming its potential as an important solution for predicting drug efficacy in rare cancers or new targets.

Seobum Seok, CEO of Lunit, stated, "With hundreds of new anticancer drug candidates being developed worldwide every year, the development of a universal immunohistochemistry AI model will greatly aid the drug development process." He added, "The importance of AI-based quantitative analysis solutions will grow even more in new anticancer drug research requiring companion diagnostic development."

This content was produced with the assistance of AI translation services.