. Author manuscript; available in PMC: 2023 Jul 15.

Published in final edited form as: IEEE J Biomed Health Inform. 2023 Jun 6;27(6):2932–2943. doi: 10.1109/JBHI.2023.3263838

TABLE V.

Evaluation Metrics (%) of Different Models on Whole Dataset

	Method	True Positive Rate	False Positive Rate	Dice Coefficient
Fully supervised baseline	U-Net	82.32 ± 5.69	0.99 ± 0.90	81.57 ± 5.71

Weakly supervised baseline	RRM	76.00 ± 16.36	5.44 ± 2.70	40.39 ± 9.01
Weakly supervised baseline	CycleGAN	55.19 ± 17.46	7.23 ± 6.44	64.14 ± 17.21

Proposed framework	U-Net	85.52 ± 5.44	1.35 ± 0.74	72.11 ± 10.12
	FCN	87.77 ± 1.01	1.49 ± 0.85	72.14 ± 8.94
	DeepLab	80.22 ± 4.18	1.45 ± 0.84	67.83 ± 1125

Ablation study	w/o Boundary masking	72.59 ± 8.46	1.37 ± 1.09	70.91 ± 16.95
	w/o Spatial regularization	77.84 1 5.67	1.21 ± 1.08	70.77 1 12.54
	Adipose seed loss + Dice loss	87.62 ± 8.79	2.24 ± 0.72	64.15 ± 9.82
	CE loss + Dice loss	74.04 ± 15.25	0.77 ± 0.30	74.99 ± 14.72

Except fully supervised method, the best and second best performance are marked in red and blue correspondingly.