Normal breast tissue classifiers assess large-scale tissue compartments with high accuracy

Abstract Cancer research emphasises early detection, yet quantitative methods for normal tissue analysis remain limited. Digitised haematoxylin and eosin (H&E)-stained slides enable computational histopathology, but artificial intelligence (AI)-based analysis of normal breast tissue (NBT) in whole slide images (WSIs) remains scarce. We curated 70 WSIs of NBTs from multiple sources and cohorts with pathologist-guided manual annotations of epithelium, stroma, and adipocytes (https://github.com/cancerbioinformatics/OASIS). We developed robust convolutional neural network (CNN)-based, patch-level classification models, named NBT-Classifiers, to tessellate and classify NBTs at different scales. Across three external cohorts, NBT-Classifiers trained on 128□x□128□µm and 256□x□256□µm patches achieved AUCs of 0.98–1.00. The model learned independent normal features different from those of precancerous and cancerous epithelium, which were further visualised using two explainable AI techniques. When integrated into an end-to-end preprocessing pipeline, NBT-Classifiers facilitate efficient downstream analysis within peri-lobular regions. NBT-Classifiers provide robust compartment-specific analytical tools and enhance our understanding of NBT appearances, which serve as valuable reference points for identifying premalignant changes and guiding early breast cancer prevention strategies. Competing Interest Statement Anita Grigoriadis, Louise J. Jones and Greg Verghese are Co-Founders of PharosAI. Salim Arslan and Pahini Pandya and employed by Panakeia Technology, UK. All other authors declare no relevant conflict of interest. Footnotes Section on Training exclusively on normal tissue enables learning of distinctive features in the normal breast updated to clarify NBT-Classifiers capture features unique to normal breast epithelium; Figure 3 revised; Supplemental files updated