Table 2.

Performance of models with and without AD on hold-out test set. Bolded values highlight the model with the best performance on a certain

	Image-Wide				Average Across Positive Images
Model Name	DSCall	Precision	Recall	F1	DSC +	IoU
UNet-AD	0.775±0.357	0.807	0.822	0.814	0.609±0.340	0.514±0.316
NestedUNet-AD	0.762±0.364	0.795	0.850	0.821	0.597±0.335	0.498±0.308
SResUNet-AD	0.754±0.362	0.793	0.832	0.812	0.575±0.321	0.469±0.290
Wang et al.-AD	0.755±0.385	0.755	0.768	0.761	0.609±0.379	0.533±0.355
SGUNet-AD	0.722±0.390	0.731	0.811	0.769	0.571±0.344	0.474±0.313
AttUNet-AD	0.776±0.355	0.804	0.846	0.824	0.623±0.334	0.527±0.313
MSU-Net-AD	0.782±0.354	0.812	0.847	0.829	0.627±0.343	0.537±0.323
UNet	0.544±0.445	0.512	0.946	0.665	0.711±0.276	0.610±0.277
NestedUNet	0.539±0.444	0.508	0.944	0.661	0.710±0.271	0.606±0.271
SResUNet	0.259±0.369	0.369	0.941	0.530	0.673±0.277	0.563±0.271
Wang et al.	0.299±0.402	0.377	0.961	0.541	0.757±0.248	0.660±0.258
SGUNet	0.442±0.435	0.458	0.920	0.611	0.621±0.289	0.506±0.273
AttUNet	0.547±0.444	0.515	0.945	0.667	0.707±0.277	0.606±0.279
MSU-Net	0.581±0.442	0.535	0.944	0.683	0.726±0.271	0.628±0.273