Table 6 Comparison of HUT and HUT-NSS noise anchor approach with the dynamic weighting (dyn) of losses or equal weighting (eq), and with the full (100%) and limited (50%, 10%, 1%) annotated dataset for downstream segmentation training.

Label count	Methods	Dice	HD95 (mm)	IOU	Precision	Recall
100%	HUT (eq)	0.598 (0.185)	14.600 (6.528)	0.471 (0.187)	0.856 (0.141)	0.503 (0.204)
100%	HUT (dyn)	0.609 (0.181)	16.591 (13.373)	0.479 (0.179)	0.701 (0.161)	0.577 (0.189)
100%	HUT-NSS (eq)	0.620 (0.184)	13.480 (6.830)	0.495 (0.185)	0.799 (0.178)	0.578 (0.186)
100%	HUT-NSS (dyn)	0.624 (0.169)	12.197 (8.780)	0.495 (0.169)	0.753 (0.152)	0.587 (0.178)
50%	HUT (eq)	0.574 (0.210)	15.017 (10.100)	0.443 (0.199)	0.703 (0.218)	0.529 (0.197)
50%	HUT (dyn)	0.591 (0.259)	13.772 (8.683)	0.476 (0.230)	0.682 (0.271)	0.563 (0.265)
50%	HUT-NSS (eq)	0.572 (0.217)	12.280 (8.082)	0.448 (0.199)	0.654 (0.211)	0.579 (0.229)
50%	HUT-NSS (dyn)	0.603 (0.252)	15.869 (18.407)	0.487 (0.235)	0.711 (0.247)	0.576 (0.251)
10%	HUT (eq)	0.486 (0.221)	24.269 (11.146)	0.363 (0.187)	0.573 (0.230)	0.492 (0.262)
10%	HUT (dyn)	0.464 (0.178)	28.418 (16.091)	0.333 (0.146)	0.666 (0.254)	0.452 (0.272)
10%	HUT-NSS (eq)	0.481 (0.233)	21.931 (14.485)	0.360 (0.197)	0.612 (0.273)	0.484 (0.301)
10%	HUT-NSS (dyn)	0.518 (0.260)	23.670 (13.740)	0.401 (0.226)	0.671 (0.280)	0.469 (0.268)
1%	HUT (eq)	0.346 (0.240)	55.780 (24.648)	0.245 (0.214)	0.731 (0.254)	0.267 (0.253)
1%	HUT (dyn)	0.336 (0.241)	58.613 (27.283)	0.238 (0.213)	0.727 (0.288)	0.254 (0.241)
1%	HUT-NSS (eq)	0.354 (0.237)	64.928 (24.281)	0.250 (0.206)	0.721 (0.278)	0.274 (0.249)
1%	HUT-NSS (dyn)	0.370 (0.240)	56.104 (27.496)	0.264 (0.208)	0.735 (0.279)	0.289 (0.252)

Quick links

Search