Identifying patients with human epidermal growth factor receptor 2 (HER2)–low and –ultralow breast cancer (BC): Use of digital, artificial intelligence (AI)–based computational algorithms to assist HER2 scoring by pathologists.

Abstract

1022 Background: Trastuzumab deruxtecan is approved in HER2-low (IHC 1+ or 2+/ISH negative) or -ultralow (IHC 0 with membrane staining in ≤10% of tumor cells) metastatic BC per DESTINY-Breast04 and -06 trials. Manual HER2 IHC scoring can be time-consuming and subjective. This retrospective, real-world study assessed manual scoring (ground truth) vs standalone AI-computational pathology-assisted (CPa) tools for scoring. Methods: Whole slide images (WSIs; N = 600, scanned with Aperio AT2) of BC samples stained with PATHWAY HER2 (4B5) assay, originally scored as HER2 IHC 0 (n = 400) or 1+ (n = 200), were rescored by 3 pathologists and using 4 CPa/AI tools in development as either IHC 0 absent membrane staining, 0 with membrane staining in ≤10% of cells, 1+, or 2+. Using 2023 ASCO/CAP guidelines, each pathologist performed 2 blinded readings per WSI; a reconciled score was used if the 2 readings differed. Manual consensus required ≥2/3 agreement. Concordance between pathologist manual consensus vs standalone CPa scoring was measured by positive and negative percentage agreement (PPA; NPA), overall percentage agreement (OPA), and Cohen κ, with review time recorded. Results: Of 600 WSIs, 586 (97.7%) had manual consensus. CPa tools were faster (Table) than manual scoring (manual median review time: 7.0 min; range, 3.0-11.5). PPA was high (≥92.5%) between manual consensus scoring and CPa tools; NPA across the tools was 79.1%, 68.9%, 67.4%, and 23.1%. Overall concordance varied across CPa tools; OPA ranged from 48.5% to 75.4% and Cohen κ from 0.26 to 0.61. Conclusions: Integrating CPa/AI tools as decision support aids for pathologists may reduce pathologist review time and augment pathologist inter-observer reproducibility, especially in HER2-ultralow identification. Refinement of CPa/AI algorithms in development may improve scoring to an even greater extent. DP tool (N = 586 a ) Review time, median (range), min PPA, b,c % (95% CI) NPA, b,d % (95% CI) OPA, b,e % (95% CI) Cohen κ e (95% CI) RV73X 2.7 (0.3-56.8) 94.9 (92.5-96.7) 79.1 (70.6-85.7) 75.4 (71.7-78.9) 0.61 (0.56-0.66) MQ52G 0.7 (0.1-8.5) 93.4 (90.9-95.4) 68.9 (60.5-76.2) 73.0 (69.3-76.6) 0.57 (0.52-0.63) KL84Q 3.1 (0.5-84.4) 92.5 (89.8-94.6) 67.4 (59.1-74.8) 69.5 (65.6-73.2) 0.54 (0.48-0.59) ZX19P 2.5 f (Not available) 99.1 (97.6-99.6) 23.1 (16.4-31.5) 48.5 (44.4-52.6) 0.26 (0.22-0.31) a Primary and metastatic samples from biopsies, effusions, fine needle aspirations, and surgical resection; due to the inbuilt pre-QC module, a few WSI outputs were not processed by RV73X and ZX19P. b Rounded to 1 decimal. c Agreement for consensus-positive cases (IHC 0 with membrane staining, 1+, or 2+). d Agreement for consensus-negative cases (IHC 0 absent membrane staining). e Agreement across all IHC scores. f Mean review time; median not available.