Qualitative evaluation scoring system.

Digital pathology enables automatic analysis of histopathological sections using artificial intelligence. Automatic evaluation could improve diagnostic efficiency and find associations between morphological features and clinical outcome. For development of such prediction models in breast cancer, identifying invasive epithelial cells, and separating these from benign epithelial cells and in situ lesions would be important. In this study, we trained an attention gated U-Net for segmentation of epithelial cells in hematoxylin and eosin stained breast cancer sections. We generated epithelial ground truths by immunohistochemistry, restaining hematoxylin and eosin sections with cytokeratin AE1/AE3, combined with pathologists’ annotations. Tissue microarrays from 839 patients, and whole slide images from two patients, were used for training and evaluation of the models. The sections were derived from four breast cancer cohorts. Tissue microarray cores from a fifth cohort of 21 patients was used as a second test set. In quantitative evaluation, mean Dice scores of 0.70, 0.79, and 0.75 were achieved for invasive epithelial cells, benign epithelial cells, and in situ lesions, respectively. In qualitative scoring (0-5) by pathologists, the best results were reached for all epithelium and invasive epithelium, with scores of 4.7 and 4.4, respectively. Scores for benign epithelium and in situ lesions were 3.7 and 2.0, respectively. The proposed model segmented epithelial cells well, but further work is needed for accurate subclassification into benign, in situ, and invasive cells.