Table 2 Cross-domain generalization on ETIS-LaribPolypDB

Method	Dice ↑	IoU ↑	BF ↑
U-Net	39.83 ± _0.92	33.59 ± _0.95	31.57 ± _1.10
U-Net++	40.12 ± _0.88	34.46 ± _0.90	32.84 ± _1.05
PraNet	62.83 ± _0.65	56.78 ± _0.68	55.12 ± _0.75
PVT	78.76 ± _0.42	70.69 ± _0.45	72.36 ± _0.55
HSNet	80.82 ± _0.38	73.44 ± _0.41	74.21 ± _0.50
ISCNet	80.42 ± _0.40	71.67 ± _0.43	73.59 ± _0.52
Polyper	86.55 ± _0.25	78.26 ± _0.28	77.47 ± _0.35
Mamba	82.52 ± _0.32	74.77 ± _0.35	75.47 ± _0.40
DDPM	78.32 ± _0.45	73.49 ± _0.48	71.18 ± _0.58
Ours	88.12 ± _0.19	79.93 ± _0.21	79.77 ± _0.26

All models are trained on Kvasir-SEG ∪ CVC-ClinicDB and evaluated on ETIS with no target-domain tuning. We report per-image mean Dice (mDice), mean IoU (mIoU), and Boundary F-measure (BF) in % (higher is better). Bold indicates the best overall. Our metrics are computed with a unified script (per-image Dice/IoU at 0.5 threshold, single-scale inference).

Quick links

Search