Table 3 Results of ablation studies with different metrics on public and private datasets.

Model	Value	AP	Precision	Recall
Public dataset
Backbone	Mean ± Std	60.5 ± 3.9%	64.2 ± 7.8%	40.5 ± 8.7%
	Median	59.2%	64.1%	38.2%
	95% CI	57.4%, 63.6%	60.7%, 68.2%	33.5%,47.4%
+DH module	Mean ± Std	66.9 ± 3.2%	67.6 ± 1.6%	59.0 ± 4.3%
	Median	67.2%	69.8%	59.7%
	95% CI	64.3%, 69.5%	65.3%, 70.9 %	55.6%,62.4%
	Effect sizes	1.791	1.267	2.171
	p value	*\(p<0.025\)	*\(p<0.025\)	*\(p<0.025\)
+ST_Encoder	Mean ± Std	68.7 ± 3.2%	69.9 ± 2.9%	61.9 ± 4%
	Median	68.5%	69.4%	63%
	95% CI	66.2%, 71.3%	67.6%,72.2 %	58.7%,65.1%
	Effect sizes	2.79	1.528	2.758
	p value	*\(p<0.025\)	*\(p<0.025\)	*\(p<0.025\)
Private dataset
Backbone	Mean ± Std	33.3 ± 2.0%	53.6 ± 3.5%	13.0 ± 5.3%
	Median	33.5%	54.0%	11.6%
	95% CI	31.7%, 34.9%	50.8%, 56.3%	8.8%,17.3%
+DH module	Mean ± Std	47.6 ± 4.4%	59.2 ± 4.5%	49.6 ± 4.4%
	Median	48.3%	59.1%	50.2%
	95% CI	44.1%, 51.2%	55.6%,62.9%	46.0%,53.1%
	Effect sizes	1.176	2.347	8.594
	p value	*\(p<0.0125\)	*\(p<0.0125\)	*\(p<0.0125\)
+ST_Encoder	Mean ± Std	49.5 ± 5.3%	56.7 ± 7.0%	52.1 ± 6.6%
	Median	49.7%	54.9%	52.3%
	95% CI	45.2%, 53.7%	51.1%,62.3%	47.8%,55.3%
	Effect sizes	3.428	1.987	4.571
	p value	*\(p<0.0125\)	*\(p<0.0125\)	*\(p<0.0125\)
+PAE module	Mean ± Std	52.6 ± 6.1%	61.3 ± 6.3%	52.8 ± 4.5%
	Median	52.6%	61.9%	52.9%
	95% CI	46.7%, 56.5%	56.3%,66.3 %	48.2%,56.4%
	Effect sizes	3.253	1.411	6.303
	p value	*\(p<0.0125\)	*\(p<0.0125\)	*\(p<0.0125\)
+T_Decoder	Mean ± Std	55.0 ± 6.4%	62.0 ± 6.4%	54.3 ± 4.2%
	Median	54.8%	62.9%	54.2%
	95% CI	49.9%, 60.1%	56.9%,67.1%	50.9%,58.6%
	Effect sizes	3.961	1.313	6.854
	p value	*\(p<0.0125\)	*\(p<0.0125\)	*\(p<0.0125\)

The metrics are shown with AP, precision and recall. The significant differences are compared between the Backbone and other modifications. For public dataset, *\(p<0.025\), Bonferroni-adjusted Wilcoxon signed rank test. For private dataset, *\(p<0.125\), Bonferroni-adjusted Wilcoxon signed rank test.

Search