Evaluation of convolutional neural networks for the detection of inter-breath-hold motion from a stack of cardiac short axis slice images

Table 3 Results of motion detection. The boldface indicates the highest score among the methods

Neural network model	Data augmentation	AUC^a	F1-score	Precision	Recall	Accuracy
Customized deep CNN with four CBR^b layers	No Aug	0.6214	0.4950	0.3646	0.7709	0.5252
Customized deep CNN with four CBR^b layers	Aug w/ flipLR	0.7073	0.5183	0.5327	0.5046	0.7168
EfficientNet-B0	No Aug	0.8656	0.6940	0.6065	0.8111	0.7841
EfficientNet-B0	Aug w/ flipLR	0.8641	0.6803	0.5573	0.8731	0.7523
MobileNet	No Aug	0.7771	0.6086	0.5366	0.7028	0.7271
MobileNet	Aug w/ flipLR	0.7709	0.5831	0.5080	0.6842	0.7047
NASNetMobile	No Aug	0.5956	0.4550	0.3890	0.5480	0.6037
NASNetMobile	Aug w/ flipLR	0.5931	0.4636	0.3662	0.6316	0.5589
ResNet50	No Aug	0.8198	0.6533	0.5755	0.7554	0.7579
ResNet50	Aug w/ flipLR	0.8242	0.6545	0.5669	0.7740	0.7533
VGG16	No Aug	0.7946	0.6163	0.6018	0.6316	0.7626
VGG16	Aug w/ flipLR	0.7884	0.5984	0.5994	0.5975	0.7579

ISSN: 1471-2342