← ResourcesPublications
DIAKWANT · HER2 · USCAP
Impact of DiaKwant HER2 (DiaDeep) Artificial Intelligence (AI) Algorithm in the Evaluation of HER2-Low and HER2-Ultralow Status in Breast Cancer: Data from a Prospective Ring Trial in Greece
Eleni (Helen) Kourea, Alexandra Lorenzo de Brionne, Florian Thomas, Triantafyllia Koletsa, Olympia Tzaida, Persefoni Xirou, Afrodite Nonni, Stylianos Malasidis, Souzana Logotheti, Yahia Salhi, Eleftherios Eleftheriadis
01 MAR 2026
Abstract
Background
HER2 0+, according to CAP HER2 evaluation, or ultralow (UL) expression, is introduced in the assessment of HER2 status in advanced breast cancer (BC), to further increase the difficulty in HER2 scoring and decrease concordance among pathologists. Herein, we aim to investigate the effect of the DiaDeep AI algorithm in the assessment of HER2-low BC cases.
Design
BC cases (n=100), originally characterized as HER2-low, were prospectively collected from 3 Pathology Laboratories. All cases were stained for HER2 with Ventana 4B5, and Agilent/Dako A0485. The HER2-stained slides were evaluated by 4 breast pathologists, according to the ASCO/CAP 2025 guidelines. DiaKwant HER2 (DiaDeep) AI algorithm, was applied to all cases for automated detection of tumor cells on whole-slide images and characterization of membrane staining, with results compared to pathologists’ evaluation. Fleiss’ κ values were calculated for each clone to explore levels of agreement among readers and readers and AI for the 0-3+ evaluation and inclusion of the 0+ category.
Results
Moderate agreement is observed between pathologists for 4B5 (Fleiss’ κ=0.507 for scores 0-3+ and κ=0.404 for 0+ inclusion). The agreement is good for A0485 in 0-3+ (κ=0.616) and moderate (κ=0.495) with 0+ inclusion. Percentage agreement of ≥3 pathologists, considered Pathologists’ Consensus (PC): for 4B5 is 84% in 0-3+ and 68% with 0+ inclusion; for A0485 is 89% in 0-3+ and 80% with 0+ inclusion. Concordance between PC and AI: for 4B5 is moderate for 0-3+ (κ=0.580) and good with 0+ inclusion (κ=0.679); for A0485 is good for 0-3+ (κ=0.703) and moderate with 0+ inclusion (κ=0.537). AI identifies more 0+ cases than PC: for 4B5 0+ cases were 8.5% by PC and 53% by AI; for A0485 0+ cases were 6.2% by PC and 36% by AI. Score 0 cases comprise: for 4B5 31.5% in PC and 7.6% in AI, concordance 76.1%; for A0485 21.3% in PC and 11.7% in AI, concordance 86.2%.
Conclusions
Agreement decreases when 0+ category is retained as a separate class for both human readers and AI. AI’s agreement closely matches human inter-observer agreement. These preliminary results suggest that AI’s performance is broadly comparable to pathologists’, with most discrepancies occurring when 0+ is treated as a distinct category. AI identifies a notably higher proportion of 0+ cases than pathologists, suggesting increased sensitivity to faint HER2 membrane staining in low-expression cases.
