Fig. 2: Regions of interest according to attention weight matrices (ANN) and eye tracking (pathologists).

A Representative image of the attention weight matrix of a BCC section. B Representative images of the ten most important tiles for the MIL-attention method of three BCC WSIs. C Representative image of the attention weight matrix of a non-BCC sample. D Representative images of the ten most important tiles for the MIL-attention method of three non-BCC WSIs. E–G Representative images of the cumulated eye traces of four board-certified pathologists on three BCC samples. H–J Representative images of the cumulated eye traces of four board-certified pathologists on three non-BCC samples. E–J Blue circles represent the artefactual retraction gaps. Red circles highlight particular focus points of eye traces. Green circles highlight epidermis, glandular structures, and hair follicles. K–N Similarity measures between a single pathologist’s eye trace and the attention weight matrix of a median performing ANN-attention model. K Heat map of Jaccard scores between pathologists and the ANN and pathologists to each other. L Scatter and bar chart of Jaccard scores between pathologists (Path-Path) and the ANN and pathologists to each other (ANN-Path; one scatter represents “path vs. path” in one image, p = 5.81 × 10^–15). M Heat map of the Sørensen–Dice coefficient between pathologists and the ANN and pathologists to each other. N Scatter and bar chart of the Sørensen–Dice coefficient between pathologists (Path-Path) and the ANN and pathologists to each other (ANN-Path; one scatter represents “path vs. path” in one image, p = 1.10 × 10⁻¹⁶). P1–P4 pathologist 1 to pathologist 4, ANN artificial neural network, BCC basal cell carcinoma, WSI whole-slide image.