HAI-score, an objective HER2 artificial intelligence method for accurate H-score estimation from IHC-stained breast cancer samples.

Abstract

1116 Background: Accurate HER2 assessment is essential for breast cancer (BC) treatment, as it directs targeted therapy decisions and predicts patient prognosis. While immunohistochemistry (IHC) is widely used, its manual scoring is susceptible to inter-observer variability. RNAscope, an RNA in situ hybridization (ISH)-based technique, has shown to have a strong correlation with HER2 protein levels and has outperformed AQUA, a high-throughput quantitative immunofluorescence imaging system, in detecting HER2-low cases. However, RNAscope is constrained by its higher cost and technical complexity compared to IHC staining assays. To address this, we propose the HAI-Score, an objective, robust, accessible, non-tissue disruptive, and fast Artificial Intelligence method for evaluating the H-score from IHC images, validated using RNAscope values. Methods: The dataset comprises 526 tissue microarray (TMA) cores for RNAscope and IHC evaluations. The dataset includes 100 commercially available BC cores (from TissueArray.Com) and 426 cores from 243 patients at MD Anderson Cancer Center. We digitized TMA cores stained with HercepTest (Dako) (S1 dataset, n=566) and Ventana Pathway 4B5 (Roche) (S2 dataset, n=580) assays. Half of the images were randomly allocated for training and the remaining half were used for validation. A computer vision algorithm detects cell membranes using a specially designed image filter based on domain knowledge. Different visual features were then extracted from these detected cell membranes, including perimeter, normalized area, Feret diameter, fractal dimension, porosity, and staining intensities. These features were used to train a neural network to predict the HAI-Score. The ground truth was defined as HER2 RNA levels measured by RNAscope. We evaluated the HAI-Score accuracy by correlating it (Pearson, R²) with RNA values and compared it to correlations from AHSQ (a state-of-the-art deep learning model), an expert pathologist, and FDA-approved HER2 IHC assays (HercepTest, Ventana PATHWAY). Results: HAI-Score yielded a correlation of 0.85 and an R 2 value of 0.711 on the testing dataset, which includes images from both the S1 and S2. This performance surpassed AHSQ, the H-score by a breast pathologist, and the scores of two FDA assays with RNA values (Table 1). Conclusions: HAI-Score provides an objective alternative to evaluate HER2 expression. It is strongly correlated with HER2 RNA levels and was superior to evaluations by an experienced breast pathologist. Following additional independent multi-site validation, HAI-Score could enable treatment personalization, optimize better surgical planning, and reduce overtreatment. Comparison of methods with RNAscope using Pearson correlation and R². Method Pearson Correlation R 2 Roche 0.58 0.33 Dako 0.76 0.57 Pathologist 0.76 0.58 AHSQ 0.83 0.69 HAI-Score 0.85 0.71