Table 8.

Comparative best round of cross-validation results of five CNN models for TB detection in the combined dataset with augmentation in nine angles

Schema	Dataset	CNN models	Best weighted						ROC-AUC
			Accuracy (%)	Recall (%)	Precision (%)	F1-Score (%)	Specificity (%)	Kappa value (%)
Without Segmentation With Augmentation	Montgomery dataset	ResNet50	75.0	80.0	75.0	77.4	69.2	49.5	85.4
		Xception	81.1	85.7	80.0	82.8	76.9	60.8	88.1
		InceptionRenNetV2	88.9	93.3	87.5	90.3	83.3	77.3	92.7
		MobileNet	81.5	92.3	75.0	82.8	71.4	63.2	88.8
		EfficientNetB3	89.3	93.3	87.5	90.3	84.6	78.4	95.3
	Shenzhen dataset	ResNet50	84.1	84.4	83.1	83.7	83.2	68.2	89.5
		Xception	85.1	84.8	84.8	84.8	85.3	70.4	90.2
		InceptionRenNetV2	91.0	90.9	90.9	90.9	91.2	82.1	94.2
		MobileNet	85.6	84.8	86.2	85.5	86.4	71.2	92.7
		EfficientNetB3	93.9	92.5	95.4	93.9	95.4	87.9	98.0
	Combination dataset	ResNet50	92.3	90.1	89.0	89.6	93.6	83.5	99.3
		Xception	91.8	89.9	87.7	88.8	92.9	82.3	99.2
		InceptionRenNetV2	93.7	90.5	92.7	91.6	95.7	86.5	99.8
		MobileNet	92.7	90.1	90.1	90.1	94.2	84.4	99.6
		EfficientNetB3	92.7	90.1	90.1	90.1	94.2	84.4	99.8
With Segmentation With Augmentation	Montgomery dataset	ResNet50	82.1	92.3	75.0	82.8	73.3	64.6	91.7
		Xception	88.9	93.3	87.5	90.3	83.3	77.3	93.3
		InceptionRenNetV2	88.9	93.3	87.5	90.3	83.3	77.3	93.8
		MobileNet	85.1	92.9	81.3	86.7	76.9	70.2	92.2
		EfficientNetB3	92.6	93.3	93.3	93.3	91.6	85.5	95.8
	Shenzhen dataset	ResNet50	90.2	89.4	90.8.	90.1	90.9	80.3	98.3
		Xception	92.4	92.3	92.3	92.3	92.5	84.8	99.2
		InceptionRenNetV2	94.7	95.3	93.8	94.6	94.1	89.4	99.6
		MobileNet	93.3	92.5	93.9	93.2	94.0	86.6	99.5
		EfficientNetB3	96.2	96.9	95.4	96.1	95.6	92.4	99.8
	Combination dataset	ResNet50	94.1	92.5	91.4	92.0	95.0	87.3	99.5
		Xception	96.0	94.0	95.1	94.5	97.1	91.3	99.6
		InceptionRenNetV2	98.2	97.5	97.5	97.5	98.5	96.1	99.9
		MobileNet	96.8	96.3	95.1	95.7	97.1	93.1	99.8
		EfficientNetB3	99.1	98.8	98.8	98.8	99.2	98.1	99.9

Best result of each set has been bolded