Table 3.

Per class and weighted average measures for method in²⁹, stand-alone CNNs and stacking ensemble of our method.

Model per class parameter	Inception V3	VGG16	VGG19	Xception	Method in29	MLP ensemble
Accuracy
Abnormal	0.948	0.792	0.922	0.883	0.931	0.974
MA	0.883	0.883	0.987	0.922	0.962	0.987
Normal	0.935	0.948	0.922	0.857	0.943	0.961
Vessel	0.896	0.961	0.987	0.974	0.946	1
Weighted average accuracy	0.913	0.91	0.958	0.916	0.945	0.982
Precision
Abnormal	0.813	0.464	1	0.619	0.952	0.928
MA	0.682	1	0.945	0.739	0.964	0.944
Normal	0.895	0.944	0.797	1	0.97	0.947
Vessel	0.95	1	0.963	1	0.964	1
Weighted average precision	0.851	0.888	0.921	0.873	0.963	0.961
Recall (sensitivity)
Abnormal	0.929	0.929	0.571	0.929	0.86	0.928
MA	0.882	0.471	1	1	0.913	1
Normal	0.85	0.85	0.95	0.45	0.907	0.9
Vessel	0.731	0.885	1	0.923	0.883	1
Weighted average recall	0.831	0.792	0.909	0.818	0.890	0.961
Specificity
Abnormal	0.952	0.762	1	0.873	0.866	0.984
MA	0.883	1	0.983	0.9	0.941	0.983
Normal	0.964	0.982	0.912	1	0.86	0.982
Vessel	0.980	1	0.980	1	0.887	1
Weighted average specificity	0.95	0.952	0.967	0.818	0.888	0.989
F1-score
Abnormal	0.867	0.619	0.727	0.743	–	0.928
MA	0.769	0.64	0.971	0.85	–	0.971
Normal	0.872	0.895	0.863	0.621	–	0.923
Vessel	0.826	0.939	0.981	0.96	–	1
Weighted average F1-score	0.833	0.803	0.902	0.808	–	0.961
Weighted average ROC-AUC score	0.981	0.97	0.99	0.989	–	0.998

Accuracy = (TP + TN)/(TP + TN + FP + FN), Precision = TP/(TP + FP), Recall(sensitivity) = TP/(TP + FN), Specificity = TN/(TN + FP), F1score = (2 × Recall × Precision)/(Recall + Precision) = (2TP)/(2TP + FP + FN), ROC-AUC score: the closer this criterion is to one, the greater the number of correctly predicted cases. On the other hand, the closer this criterion is to zero, the greater the number of incorrectly predicted cases.