Table 2.

Impact of region of interest detection on diagnostic efficiency of convex probe endobronchial ultrasound multimodal images

Modes	Five-fold cross-validation
	1		2		3		4		5		Average
	AUC	P	AUC	P	AUC	P	AUC	P	AUC	P	Average AUC 95% CI	P
G w/o ROI	0.6725	0.7843	0.6850	0.0847	0.5661	0.1206	0.6705	0.3635	0.6011	0.0938	0.6390 (0.5740-0.7041)	0.91984
G w/ROI	0.6814		0.6181		0.6297		0.6212		0.6586		0.6418 (0.6079-0.6757)
F w/o ROI	0.6363	0.3089	0.6541	0.2958	0.6574	0.9650	0.6866	0.6029	0.5201	0.0253	0.6309 (0.5508-0.7110)	0.62530
F w/ROI	0.5841		0.7053		0.6552		0.6603		0.6386		0.6487 (0.5944-0.7030)
E w/o ROI	0.9548	0.0625	0.9488	0.4774	0.9456	0.2956	0.9579	0.3211	0.9447	0.0195	0.9503 (0.9432-0.9575)	0.13900
E w/ROI	0.9806		0.9348		0.9626		0.9727		0.9792		0.9660 (0.9426-0.9894)

Bold indicates P<0.05. The first row shows the performance of the DL model which is trained and tested with cropped images, while the second row corresponds to the models trained and tested with ROI. For each fold, AUCs are calculated for each mode with and without ROI, and P value is obtained using the Delong test between the two ROCs. For average AUCs of five-folds, P value in the last column is obtained with paired-samples t-test. G: Gray scale; F: Blood flow Doppler; E: Elastography; DL: Deep learning; ROI: Region of interest; w/o ROI: Without ROI; w/ROI: With ROI; AUCs: Area under the curves; CI: Confidence interval