EVIDENCE BASED SELECTION OF HYPERPARAMETERS FOR EPITHELIAL SEGMENTATION IN WHOLE SLIDE IMMUNOHISTOCHEMISTRY IMAGES: FROM NORMAL MUCOSA TO COLORECTAL CANCER

Abstract

Background. Accurate delineation of the epithelial compartment is a critical prerequisite for reproducible quantitative analysis in digital pathology, particularly in colorectal cancer (CRC) where malignant glands determine most prognostic and biomarker-driven readouts. However, robust and transferable epithelial segmentation strategies for immunohistochemistry (IHC) whole-slide images remain insufficiently standardized.Methods. We implemented a two-stage experimental design to optimize and validate deep learning-based epithelial segmentation on DAB hematoxylin IHC images. In the first stage, U-Net hyperparameters were systematically screened on normal colorectal mucosa using a controlled factor-wise approach, evaluating encoder depth, activation function, and optimizer. Model performance was assessed using aggregated pixel-level metrics with emphasis on overlap-based measures. In the second stage, the selected configuration was transferred without modification to primary CRC tissue and used to train and evaluate two independent binary segmentation models for CD163 and CD80 IHC datasets.Results. Factor screening identified a deeper U-Net architecture with ReLU activation and RMSprop optimization as the most reliable configuration. When applied to CRC tissue, the CD163 model achieved high segmentation performance with strong agreement between predicted and ground-truth epithelial areas, while the CD80 model demonstrated consistently superior overlap metrics and near-perfect agreement in epithelial area estimation. The optimized configuration showed stable performance across markers, indicating robustness to staining heterogeneity and class imbalance.Conclusions. A structured, design-of-experiments-guided approach to U-Net parameter selection enables robust and transferable epithelial segmentation across different IHC markers in CRC. This strategy supports reproducible compartment-aware quantitative analysis and facilitates downstream biomarker assessment in digital pathology workflows.