Classification of chest X-ray images by incorporation of medical domain knowledge into operation branch networks

BMC Medical Imaging

Table 5 Results obtained using earlier methods with the NIH 14 dataset

Method	Atel	Card	Effu	Infi	Mass	Nodu	Pne1	Pne2	Cons	Edem	Emph	Fibr	PT	Hern	Mean
Wang et al. [22]	0.716	0.807	0.784	0.609	0.706	0.671	0.633	0.806	0.708	0.835	0.815	0.769	0.708	0.767	0.738
CheXNet [1]	0.809	0.925	0.864	0.735	0.868	0.780	0.768	0.889	0.790	0.888	0.937	0.805	0.806	0.916	0.841
AG-CNN [10]	0.853	0.939	0.903	0.754	0.902	0.828	0.774	0.921	0.842	0.924	0.932	0.864	0.837	0.921	0.871
Triple Attention [34]	0.779	0.895	0.836	0.710	0.834	0.777	0.737	0.878	0.759	0.855	0.933	0.838	0.791	0.938	0.826
Anatomy X-Net [16]	0.815	0.908	0.880	0.705	0.855	0.793	0.775	0.874	0.810	0.896	0.923	0.832	0.786	0.962	0.844
ResNet50 (Backbone)	Conventional	0.746	0.874	0.809	0.694	0.762	0.709	0.686	0.835	0.733	0.837	0.869	0.793	0.745	0.913	0.786
	ABN	0.746	0.874	0.813	0.686	0.777	0.713	0.679	0.838	0.731	0.838	0.873	0.801	0.752	0.914	0.788
	OBN1	0.745	0.884	0.812	0.691	0.773	0.708	0.674	0.837	0.729	0.834	0.867	0.811	0.751	0.908	0.787
	OBN2	0.747	0.881	0.805	0.695	0.775	0.705	0.689	0.841	0.723	0.839	0.873	0.796	0.750	0.884	0.786
DenseNet121 (Backbone)	Conventional	0.744	0.879	0.814	0.683	0.791	0.701	0.694	0.822	0.731	0.833	0.826	0.793	0.750	0.883	0.782
	ABN	0.745	0.885	0.814	0.698	0.783	0.718	0.692	0.842	0.735	0.836	0.884	0.813	0.759	0.892	0.792
	OBN1	0.743	0.885	0.811	0.692	0.764	0.710	0.684	0.848	0.728	0.837	0.881	0.808	0.756	0.905	0.789
	OBN2	0.742	0.883	0.810	0.697	0.767	0.716	0.682	0.851	0.733	0.838	0.877	0.795	0.753	0.917	0.790

The AUC of each class and the average AUC of 14 diseases are shown. These diseases for the NIH 14 dataset are atelectasis (Atel), cardiomegaly (Card), effusion (Effu), infiltration (Infi), mass (Mass), nodule (Nodu), pneumonia (Pne1), pneumothorax (Pne2), consolidation (Cons), edema (Edem), emphysema (Emph), fibrosis (Fibr), pleural thickening (P.T.), hernia (Hern), with ResNet50 and DenseNet121 used as backbone approaches. ABN, attention branch network; OBN1, operation branch network using a weight map with a convex hull on mask images of the lung field; OBN2, operation branch network using weight maps with combined mask images of the lung field and heart

ISSN: 1471-2342