Abstract 5388: A unified computational pathology method to quantify HER2 expression from raw IHC and IF images in breast cancer

Background: The development of automated quantitative methods to measure biomarker expression such as HER2 aims at reducing the subjective variability of pathologist-performed biomarker assessment of immunohistochemically (IHC) stained tissue slides. As an example, the deep learning-based Quantitative Continuous Scoring (QCS) algorithm [1] enables the objective measurement of biomarker expression based on the detection of individual tumor cells in the tumor regions and - for each cell, on the instance segmentation of its nucleus, cytoplasm, and membrane compartments. Methods: In order to extend the QCS image analysis to similarly analyze tissue slides stained with immunofluorescence (IF), we re-trained the fully supervised deep-learning models, adjusted the normalization of images on tissue controls to account for variability between different staining batches, and finally used the normalized signal in IF instead of the Optical Density (OD) signal in IHC as a 8-bit grayscale image on which the biomarker expression is estimated. We performed QCS on HER2 IF images (HER2 clone 29D8 [CST], imaged with Vectra [Akoya] in parallel with HER2 IHC clone 4B5 [Roche Tissue Diagnostics]) performed on 26 primary and metastatic breast cancer samples representing the full range of HER2 expression, from null to highly overexpressed. Results: Our analysis demonstrated that the QCS-based scoring on IHC-HER2 images correlates with the QCS-based scoring on IF-HER2 images. We observed a Pearson correlation of R=0.92 between the median membrane OD in IHC and the median normalized membrane signal in IF. Defining a positive cell as having an estimated membrane expression higher than a given so-called positivity threshold, we found a median Pearson correlation of R=0.85 between the percentage of positive cells detected in IHC and the percentage of positive cells detected in IF for increasing values of positivity thresholds. Correlation of the QCS median normalized membrane signal in IF was R=0.91 with mRNA (ERBB2 transcript levels [Nano String]) and R=0.88 with IHC-based H-scores, against R=0.83 and R=0.87 respectively for the QCS median membrane OD signal in IHC. Conclusion: We describe the extension of a computational pathology-based approach for biomarker quantification in IHC to IF stained tissue slides. The consistency of the image analysis method translates into the consistency of the measurements in the two imaging methods. The use of IF could enable the improved quantificat