Table 1.

Description of datasets used in the study.

Dataset	Unique Cases	Slides	Description
LocalSentinel	161	161	Retrospectively collected lymph node slides from sentinel node procedures consisting of 107 negative and 54 positive slides. One slide per unique patient case included. Cytokeratin immunohistochemically stained sections were available (AE1/AE3) and were used to aid in production of detailed ground-truth annotations. (Subset of AIDA BRLN dataset.)
LocalAxillary	48	57	Retrospectively collected lymph node slides from axillary dissection procedures consisting of 24 negative and 24 positive slides from unique cases. This dataset was further enriched with nine extra negative slides, overlapping with other cases, which contained only extra-nodal tissue and featured examples of fat necrosis and foreign body tissue reactions. (Subset of AIDA BRLN dataset.)
LocalNegativeAxillary	24	259	Retrospectively collected lymph node cases from axillary dissection procedures that were signed out as negative during initial clinical diagnostic assessment. All slides in the cases were included except for slides (n = 33) already included in LocalSentinel or LocalAxillary to avoid overlapping slides. One section per block, except in four cases 1–3 additional sections per block were presented in the clinical archive and included in this set. (Subset of AIDA BRLN dataset.)
CAMELYON16	399	399	Retrospectively collected sentinel lymph node slides from two hospitals in the Netherlands. Pre-generated data splits by the CAMELYON organizers were retained in this study. Of the 160 positive slides, detailed annotations were available for 140 slides, whereas 20 slides were only partially annotated e.g., slides that contained two consecutive sections of the same tissue or slides that contained out-of-focus tumor regions, and only the delineated regions containing tumors were used in these slides during training.
CAMELYON17	200	1000 (344 used)	Retrospectively collected sentinel lymph node slides from five hospitals in the Netherlands. A total of 1000 slides. Only a subset of 50 positive slides contained detailed annotations, while the remaining positive slides had been given a slide label according to the TNM staging system. Of the available 50 annotated positive slides, only 34 slides contained either micro- or macro-metastases and were included in this study, the remaining 16 slides were labelled as isolated tumor cells (ITC) and were excluded. 50 randomly selected negative slides from the CAMELYON17 training set and 260 negative slides in the CAMELYON17 test set were allocated to training and testing sets, respectively, in this study. In total 344 slides from CAMELYON17 were used in this study.