Application of artificial intelligence-driven endoscopic screening and diagnosis of gastric cancer

Table 1 Summary of artificial intelligence applications in predicting Helicobacter pylori infection

Ref.	Endoscopic modality	Training dataset	Validation dataset	Accuracy	Sensitivity	Specificity	PPV
Huang et al[78], 2004	WLI	30 patients	74 patients	85.1 (avg)1	78.8 (avg)	90.2 (avg)	-
Shichijo et al[79], 2017	WLI	32208 images, 1768 patients	11481 images, 397 patients	87.7	88.9	87.4	-
Itoh et al[81], 2018	WLI	149 images, 139 patients	30 images, 30 patients	-	86.7	86.7	-
Nakashima et al[84], 2018	WLI, BLI and LCI	162 patients	60 patients	-	96.7	-	-
Shichijo et al[80], 2019	WLI	98564 images, 4494 patients	23699 images, 847 patients	Infected: 66.0; post-eradication: 86.0	-	-	-
Zheng et al[82], 2019	WLI	11729 images, 1507 patients	3755 images, 452 patients	84.5	81.4	90.1	-
Zhu et al[100], 2019	WLI	790 images	203 images	89.2	76.5	95.6	89.7
Nakashima et al[85], 2020	WLI, BLI and LCI	12887 images, 395 patients	120 patients	80.0 (avg)2	61.3 (avg)	89.4 (avg)	74.7 (avg)

¹Histological characteristics were assessed for the various antrum, body and cardia locations.

²White light imaging and linked color imaging-based images were both analyzed, with the linked color imaging obtaining significantly higher accuracy, sensitivity, specificity and positive predictive value. BLI: Blue laser imaging; LCI: Linked color imaging; PPV: Positive predictive value; WLI: White light imaging.

Table 2 Summary of artificial intelligence applications in prediction of invasion depth and differentiation of cancerous areas from noncancerous areas

Ref.	Application	Endoscopic modality	Training dataset	Validation dataset	Accuracy	Sensitivity	Specificity	PPV	NPV
Kubota et al[98], 2012	Prediction of invasion depth	WLI	344 patients, 902 images	-	77.2 (T1)	-	-	80.1 (T1)
Miyaki et al[137], 2013	Differentiation of cancerous areas from noncancerous areas	WLI and magnified FICE	493 images	46 images	85.9	84.8	87.0	86.7	85.1
Hirasawa et al[99], 2018	Differentiation of cancerous areas from noncancerous areas	WLI	13584 images	2296 images, 69 patients	92.2	92.2	-	30.6	-
Kanesaka et al[138], 2018	Detection of EGC	Magnified NBI	126 images	81 images	96.3	96.7	95.0	98.3	-
Horiuchi et al[103], 2020	Differentiation of EGC from gastritis	Magnified NBI	2570 images	258 images	85.3	95.4	71.0	82.3	91.7
Yoon et al[101], 2019	Detection of EGC and prediction of EGC invasion depth	WLI	11686 images, 800 patients	-		79.2	77.8	79.3	77.7
Horiuchi et al[105], 2020	Detection of EGC	Magnified NBI	2570 images	174 videos, 82 patients	85.1	87.4	82.8	83.5	86.7
Li et al[139], 2020	Differentiation of EGC from noncancerous lesions	Magnified NBI	2088 images	342 images	90.9	91.2	90.6	90.6	91.2
Nagao et al[102], 2020	Prediction of invasion depth	WLI, nonmagnifying NBI and indigo-carmine dye contrast imaging (Indigo)	16557 images, 1084 patients	-	94.4	89.2	98.7	98.3	91.7
Namikawa et al[104], 2020	Differentiation of cancerous areas from noncancerous areas	WLI, nonmagnifying NBI and indigo-carmine dye contrast imaging (Indigo)	18410 images	1459 images	95.9	99.0	93.3	92.5	-

EGC: Early gastric cancers; FICE: Flexible spectral imaging color enhancement; NBI: Narrow-band imaging; NPV: Negative predictive value; PPV: Positive predictive value; WLI: White light imaging.