. Author manuscript; available in PMC: 2022 Jan 1.

Published in final edited form as: Med Image Anal. 2020 Oct 7;67:101841. doi: 10.1016/j.media.2020.101841

Table 1.

Class distribution statistics: The top portion reports the distribution of labels for both scenarios. In Scenario 1, we were able to balance distribution across training and test sets to within 1% (stratified cross-validation). Class distributions in training and test sets are explicitly given for Scenario 2. In the bottom portion, we report on the size of the datasets in terms of both images and labeled pixels, as well as on the overall fraction of pixels that were labeled.

	Scenario 1	Scenario 2
	Whole Dataset	Europe (Train)	US (Test)
Background	17% (83.7M)	19% (65.3M)	13% (18.4M)
Artifact	19% (92.6M)	20% (67.7M)	18% (24.5M)
Mesh	20% (97.5M)	21% (73.7M)	17% (23.8M)
Nest	5% (25.6M)	6% (19.0M)	5% (6.6M)
Ring	28% (136.6M)	23% (78.8M)	40% (57.8M)
Aspecific	10% (50.5M)	12% (40.3M)	7% (10.1M)

labeled Pixels	57% (486.6M)	51% (344.7M)	60% (141.8M)
# labeled Images	117	86	31