Table 1.

Comparison of Dice scores for lesion segmentation between various weakly supervised networks and ours. Note that our evaluation only used 10 of 20 images in the CoronaCases dataset due to lack of sufficient coronal resolution on the other 10. Other networks in this chart trained on axial images and were therefore able to make use of all 20.

Method	Number of manual segmentations used in training	Dice score
ActiveLearning	16	52.4
CoSinGan	2	57.8
GASNet	1	70.3
Ours	0	55.9