CXR-RefineDet: Single-Shot Refinement Neural Network for Chest X-Ray Radiograph Based on Multiple Lesions Detection

Table 3

Detection results of different methods on VinDr-CXR test set.

Method	Backbone	Input size	mAP	Inference speed (fps) w/o reparam	Params (M)
RetinaNet	ResNet-50	512 × 512	0.1269	3.7	36.37
	ResNet-101		0.1494	3.5	55.37
Faster RCNN	ResNet-50	512 × 512	0.1597	3.7	41.19
	ResNet-101		0.1569	3.5	60.18
GA-Faster RCNN	ResNet-50	512 × 512	0.1464	3.3	41.78
Cascade RCNN	ResNet-50	512 × 512	0.1512	3.3	68.97
	ResNet-50-DCN		0.1613	3.2	69.55
	ResNet-101		0.1631	3.1	87.96
	ResNet-101-DCN		0.1560	3.0	89.24
VFNet	ResNet-50	512 × 512	0.1413	3.6	32.51
	ResNet-101		0.1505	3.4	51.51
ATSS	ResNet-50	512 × 512	0.1478	3.7	31.92
	ResNet-101		0.1538	3.5	50.91
RefineDet	VGG-16	320 × 320	0.1149	10.8	—
RefineDet	VGG-16	512 × 512	0.1173	9.9	33.51
CXR-RefineDet (ours)	RRNet	320 × 320	0.1392	9.6	—
CXR-RefineDet (ours)	RRNet	512 × 512	0.1618	6.8	49.76