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

BMC Medical Imaging

Table 4 Heart failure classification results from ten-fold cross-validation

Backbone	Model	Accuracy [%]	Sensitivity [%]	Specificity [%]	AUC
ResNet50	Conventional	60.26 ± 3.64	37.59 ± 15.63	82.92 ± 10.41	0.68 ± 0.03
	ABN	59.20 ± 5.81	81.09 ± 17.98	67.31 ± 10.46	0.67 ± 0.06
	OBN1	62.49 ± 3.90	58.75 ± 16.44	66.24 ± 13.27	0.70 ± 0.04
	OBN2	64.38 ± 2.79	58.32 ± 8.62	70.44 ± 7.31	0.69 ± 0.03
	OBN3	61.99 ± 3.66	54.87 ± 11.32	69.11 ± 6.24	0.68 ± 0.04
DenseNet121	Conventional	61.29 ± 5.45	39.46 ± 16.01	83.12 ± 8.54	0.70 ± 0.06
	ABN	65.24 ± 5.09	61.00 ± 15.36	69.49 ± 11.55	0.70 ± 0.05
	OBN1	63.79 ± 4.30	55.65 ± 17.19	71.93 ± 11.20	0.69 ± 0.03
	OBN2	66.54 ± 5.21	64.40 ± 10.63	68.67 ± 5.07	0.71 ± 0.05
	OBN3	63.80 ± 3.47	58.03 ± 16.14	69.57 ± 15.25	0.70 ± 0.05

Bolded numbers indicate the highest score
ResNet50 and DenseNet121 are used as backbone approaches: ABN, attention branch network; OBN1, operation branch network using weight map with a convex hull on mask images of lung field; OBN2, operation branch network using weight maps with combined mask images of the lung field and heart; OBN3, operation branch network using weight maps manually masked to include the heart according to the doctor's support

ISSN: 1471-2342