Generalizing AI-Driven Assessment of Immunohistochemistry Across Immunostains and Cancer Types: A Universal Mmunohistochemistry Analyzer

Biagio Brattoli, Mohammad Mostafavi, Taebum Lee, Wonkyung Jung, Jeongun Ryu, Seonwook Park, Jongchan Park, Sérgio Pereira, Seunghwan Shin, Sangjoon Choi, Hyojin Kim, Donggeun Yoo, Siraj M. Ali, Kyunghyun Paeng, Chan-Young Ock, Soo Ick Cho, Seokhwi Kim

Abstract:

Background: Despite the advancements in novel methodologies, immunohistochemistry (IHC) remains the most widely utilized ancillary test for histopathologic and companion diagnostics across various targeted therapies. However, the objective assessment of IHC has long posed challenges. Artificial intelligence (AI) has emerged as a potential solution; however, its development process demands extensive training specific to each cancer and IHC type, limiting its versatility.

Methods: We developed a Universal IHC (UIHC) analyzer, an AI model tailored to interpret IHC images irrespective of tumor or IHC types, utilizing training datasets from multiple cancer types stained for PD-L1 and/or HER2.

Findings: This multi-cohort trained model outperforms conventional single-cohort trained models in interpreting unseen IHCs (Kappa score 0.578 vs. up to 0.509). The UIHC model consistently demonstrates superior performance across a wide variety of cutoff values for positive staining. Qualitative analysis on the learned representation reveals that UIHC effectively separates and clusters patches based on expression levels. The UIHC model enables quantitative assessment of c-MET expression in the real-world NSCLC cases harboring MET mutations.

Interpretation: These findings highlight the potential of the UIHC model as a universal analyzer for IHC images, representing a significant advancement in AI application in pathology amidst the era of personalized medicine and accumulating novel biomarkers.