Lizard contains subsets of other datasets

[ClaudiaWinklmayr]

Quotes from Lizard publication

• "To generate our final dataset, we considered data from 6 different sources: GlaS [34], CRAG [17], CoNSeP [20], DigestPath, PanNuke [15] and TCGA [21]. " Lizard-paper
• "For both CoNSeP and PanNuke, we only utilised the images and not the associated annotations from the original datasets to ensure that the same label generation pipeline was used on all input data" Lizard-paper

Caution when training

• when training/testing with the prepared datasets, we need to be aware that they are not mutually exclusive but have some degree of overlap
• In particular when using CoNSep and PanNuke, we need to be aware that they are part of Lizard but their annotations in Lizard may differ from the original annotations.

Potential for leakage

• Phikon-v2 is (among others) trained on TCGA, which is also a part of Lizard. See this issue