. Author manuscript; available in PMC: 2023 Aug 1.

Published in final edited form as: IEEE Trans Med Imaging. 2022 Aug 1;41(8):1990–2003. doi: 10.1109/TMI.2022.3153322

TABLE VI:

The results of cross-dataset evaluation. The AUROC and AUPRC results of various models trained on CheXpert dataset and evaluated on MIMIC-CXR dataset. The upper part is the AUROC results and the lower part is the AUPRC results.

		EC	Card	AO	LL	Edem	Cons	Pneu1	Atel	Pneu2	PE	PO	Frac	SD	mean

AUROC	A-GCN-PPS	0.567	0.770	0.708	0.661	0.884	0.754	0.603	0.764	0.752	0.895	0.701	0.610	0.839	0.731
	A-GCN-APS	0.552	0.754	0.686	0.661	0.852	0.720	0.604	0.743	0.717	0.864	0.685	0.587	0.797	0.709
	AlexNet	0.563	0.758	0.704	0.655	0.880	0.753	0.619	0.760	0.736	0.888	0.701	0.604	0.835	0.727

	R-GCN-PPS	0.530	0.768	0.721	0.679	0.885	0.772	0.665	0.761	0.771	0.902	0.757	0.636	0.842	0.745
	R-GCN-APS	0.558	0.762	0.702	0.663	0.855	0.730	0.629	0.730	0.742	0.880	0.721	0.615	0.796	0.722
	ResNet50	0.527	0.713	0.671	0.609	0.842	0.690	0.606	0.702	0.700	0.857	0.655	0.573	0.781	0.687

	V-GCN-PPS	0.576	0.768	0.721	0.673	0.887	0.761	0.638	0.758	0.785	0.904	0.743	0.641	0.842	0.746
	V-GCN-APS	0.547	0.763	0.689	0.662	0.854	0.720	0.610	0.726	0.746	0.875	0.713	0.614	0.806	0.717
	VGGNet16BN	0.520	0.709	0.690	0.634	0.854	0.687	0.596	0.701	0.718	0.878	0.672	0.582	0.809	0.696

	p-val (RM-ANOVA)		^*		^*	^*			^*	^*	^*			^*
	PPS >APS		^*	^**	^*	^***	^***		^***	^***	^***	^**	^***	^***	^***
	PPS >base		^*	^*	^*	^*	^*		^*	^*	^*	^*	^*	^*	^*

AUPRC	A-GCN-PPS	0.042	0.414	0.353	0.064	0.459	0.110	0.123	0.373	0.154	0.688	0.023	0.031	0.572	0.262
	A-GCN-APS	0.039	0.395	0.333	0.064	0.389	0.094	0.119	0.352	0.124	0.625	0.026	0.028	0.491	0.237
	AlexNet	0.042	0.402	0.351	0.060	0.453	0.108	0.134	0.370	0.137	0.682	0.026	0.030	0.565	0.259

	R-GCN-PPS	0.038	0.418	0.373	0.072	0.467	0.112	0.147	0.366	0.208	0.711	0.034	0.036	0.554	0.272
	R-GCN-APS	0.045	0.405	0.347	0.068	0.407	0.099	0.130	0.339	0.168	0.660	0.029	0.033	0.492	0.248
	ResNet50	0.031	0.360	0.315	0.050	0.388	0.084	0.117	0.313	0.131	0.628	0.021	0.028	0.475	0.226

	V-GCN-PPS	0.043	0.417	0.372	0.068	0.472	0.110	0.142	0.370	0.201	0.711	0.032	0.036	0.558	0.272
	V-GCN-APS	0.037	0.404	0.336	0.063	0.406	0.092	0.123	0.338	0.170	0.647	0.030	0.033	0.507	0.245
	VGGNet16BN	0.032	0.359	0.334	0.056	0.409	0.086	0.118	0.312	0.138	0.671	0.023	0.029	0.517	0.237

	p-val (RM-ANOVA)		^*		^*	^*				^**	^*		^*	^*
	PPS >APS		^***	^**	^*	^***	^***	^*	^***	^***	^***		^***	^***	^***
	PPS >base		^*	^*	^*	^*	^*		^*	^*	^*		^*	^*	^*

*, ** and *** denote the significance levels of 0.1, 0.05 and 0.01, respectively. For each disease, the best results are bolded The red text means our ImageGCN can perform better than the corresponding two baseline models.

EC: Enlarged Cardiomediastinum; Card: Cardiomegaly; AO: Airspace Opacity; LL: Lung Lesion; Edem: Edema; Cons: Consolidation; Pneul: Pneumonia; Atel: Atelectasis; Pneu2: Pneumothorax; PE: Pleural Effusion; PO: Pleural Other; Frac: Fracture; SD: Support Devices