Abstract 193: A deep learning workflow to quantify the H-Score on immunohistochemistry images

Abstract

Abstract Quantification of protein expression in IHC is typically assessed with H-scores, calculated by a pathologist on a scale from 0 to 300 based on the intensity of carcinoma cells and their relative proportion.The subjectivity of the methodology is an inherent risk of inconsistency for inter- intra- assay. Here, we developed a semi-automated workflow based on deep learning to calculate the H-score on IHC slides for the CDH17 protein, a component of the gastrointestinal (GI) system. CDH17 membranous expression makes its quantification particularly challenging and prone to variability. This workflow was validated on 112 samples of GI tumors (primary as well as metastatic), reaching an H-score correlation score R2=0.92 (p-value below 10-5) and an overall median absolute error=17.5 compared to a pathologist's evaluation. In our method, tumor regions on each slide image were first annotated by a pathologist, containing clusters of epithelial carcinoma mixed with non-tumor tissue. A segmentation model was trained to outline contours of marked carcinoma clusters, later classified as a low, moderate or strong expression based on optical density distribution. Finally, all cluster results were aggregated at slide level to get a final H-score. The method was evaluated on 112 tumor sections including the following cancer histotypes: pancreatic (N=20), gastric (N=18), esophageal (N=20), cholangiocarcinoma (N=20) and colorectal (N=34). A U-Net segmentation model with a ResNet101 encoder was trained on 303 annotated regions of interest (ROIs) of size 240x240 µm2 taken from 6 images, scanned at 20x. The model was taught to segment CDH17- and CDH17+ carcinoma clusters with 80% of the annotated ROIs. Once trained, the model reached a mIoU=0.74 and mDice=0.85 on a previously unseen evaluation set compound of the remaining 20% ROIs. The automatic method was then validated to attribute a score to each segmented CDH17+ cluster. DAB expressions were extracted for each cluster using color deconvolution and the best classification performances were obtained using the 90th percentile of the DAB distribution (Kappa score=0.74) compared to pathologist scoring. Finally, the whole workflow was launched on the entire image dataset to get predicted H-scores and compare them to the pathologist ones. The workflow reached a correlation R2=0.92 and a median absolute error=17.5 over all tumor locations. Our workflow reproduces the pathologist H-Score and shows a consistent performance on low- and high- range H-Scores (i.e. close to 0 and superior to 200), across all tumor locations of the study. This methodology fosters robustness and reproducibility of H-score quantification results for the CDH17 biomarker. Our workflow can straightforwardly be applied to other histological biomarkers, aiming to increase the rigorousness of its validation for future application in patient selection assays. Citation Format: Paul Klein, Thomas Le Meur, Roberto Fiorelli, Claudia Buehnemann, Aurelie Auguste, Sylvain Berlemont. A deep learning workflow to quantify the H-Score on immunohistochemistry images [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 193.