Semisupervised Medical Image Segmentation through Prototype-Based Mutual Consistency Learning

Table 4

Quantitative comparison between proposed PMCL and other semisupervised methods on Kvasir-SEG under 10% and 20% labeled data.

Method	Labeled/unlabeled	DSC (%)	JI (%)	SE (%)	AC (%)	95HD (mm)	ASD (mm)
Fully supervised	800/0	81.79	73.16	84.62	95.13	77.60	23.53
Supervised-only	80/0	73.04	62.32	80.32	92.39	117.23	46.83
MT [5]	80/720	74.09	63.31	82.33	92.90	115.48	41.87
DAN [10]	80/720	75.29	65.14	81.33	92.88	106.65	39.26
EM [49]	80/720	74.73	64.38	83.75	92.69	121.12	42.55
UAMT [8]	80/720	74.66	64.32	81.61	92.68	112.18	38.61
ICT [50]	80/720	74.58	64.49	80.83	93.02	100.03	37.46
PMCL (ours)	80/720	76.02	66.40	81.53	93.56	91.03	30.96
Supervised-only	160/640	77.94	69.26	83.12	94.09	94.98	32.04
MT [5]	160/640	78.14	69.52	77.29	94.40	77.65	21.14
DAN [10]	160/640	78.38	70.04	78.79	94.48	81.64	20.31
EM [49]	160/640	78.59	69.47	83.23	94.28	91.34	31.94
UAMT [8]	160/640	78.42	69.96	80.30	94.56	78.85	23.26
ICT [50]	160/640	78.78	70.55	78.90	94.48	70.33	19.71
PMCL (ours)	160/640	79.98	70.92	84.24	94.63	97.83	31.01

The bold values suggest the best performance compared to other state-of-the-art methods.