Table 2.

Performances of different methods in prostate gland segmentation in terms of pixel-based metrics

Method	Accuracy	Precision	Recall	Dice
Farjam et al.25,a	0.6378 ± 0.1586	0.7183 ± 0.3034	0.4372 ± 0.1736	0.5070 ± 0.2059
Naik et al.26,a	0.7402 ± 0.1151	0.7958 ± 0.2021	0.5819 ± 0.2275	0.6357 ± 0.2105
Peng et al.27,a	0.7957 ± 0.1535	0.6508 ± 0.2568	0.9305 ± 0.1124	0.7334 ± 0.2198
Nguyen et al.28,a	0.7703 ± 0.1632	0.8260 ± 0.1588	0.7041 ± 0.2998	0.7145 ± 0.2556
Singh et al.29,a	0.6734 ± 0.1247	0.9001 ± 0.1743	0.3869 ± 0.2493	0.4931 ± 0.2557
Ren et al.30,b	0.8576 ± 0.1139	0.8199 ± 0.1638	0.8861 ± 0.1673	0.8308 ± 0.1495
Xu et al.31,b	0.8250 ± 0.1106	0.7407 ± 0.1597	0.9273 ± 0.1079	0.8079 ± 0.1264
Salvi et al.24,b	0.9325 ± 0.0684	0.8897 ± 0.1359	0.9356 ± 0.0964	0.9016 ± 0.1087
Mask R-CNN19,b,c	0.9410 ± 0.0010	0.9002 ± 0.0026	0.9468 ± 0.0011	0.9229 ± 0.0015

The performances were on the hold-out test set of Salvi et al.²⁴ Note that accuracy values were the balanced accuracy values as in Salvi et al.,²⁴ and all performance values except the one for the Mask R-CNN model were collected from Salvi et al.²⁴

^aTraditional image processing or machine-learning-based methods.

^bDeep-learning-based methods.

^cStandard deviations were calculated using bootstrapping.²²