Table 9.

Impact of sequential semi-supervised learning strategy on classification performance.

Method	s-SSL	Accuracy	Precision	Recall	$F_1$-Score	AUC	p-value
R18+RMTL	✗	0.577 ± 0.080	0.645 ± 0.015	0.636 ± 0.024	0.575 ± 0.082	0.719 ± 0.102	***
	✓	0.842 ± 0.023	0.859 ± 0.027	0.797 ± 0.029	0.814 ± 0.029	0.916 ± 0.012	–
R50+RMTL	✗	0.710 ± 0.085	0.739 ± 0.179	0.674 ± 0.093	0.662 ± 0.029	0.789 ± 0.092	**
	✓	0.825 ± 0.014	0.833 ± 0.016	0.780 ± 0.020	0.795 ± 0.018	0.915 ± 0.016	–
R18+LA-Net	✗	0.759 ± 0.071	0.750 ± 0.043	0.766 ± 0.038	0.750 ± 0.062	0.829 ± 0.016	***
	✓	0.865 ± 0.006	0.857 ± 0.011	0.860 ± 0.011	0.858 ± 0.011	0.924 ± 0.010	–
R50+LA-Net	✗	0.827 ± 0.090	0.816 ± 0.097	0.816 ± 0.055	0.814 ± 0.081	0.901 ± 0.030	*
	✓	0.888 ± 0.008	0.886 ± 0.010	0.869 ± 0.011	0.876 ± 0.009	0.947 ± 0.009	–

Notes: This experiment was conducted on the partly-labeled MBUD dataset. The symbols ✓ and ✗show whether the method is trained with or without our sequential semi-supervised learning strategy. ResNet50 was used as FEX, and the input image size was 256 × 256 pixels. The outcomes are represented as mean value ±95% confidence interval. The null hypothesis and meaning of p-values remain consistent with the description in Table 3.