Artificial intelligence technique in detection of early esophageal cancer

Table 1 Application of artificial intelligence in endoscopic detection of early esophageal cancer

Ref.	Modality	AI technique	No. of images/cases in training dataset (positive/negative)	No. of images/cases in test dataset (positive/negative)	Results
van der Sommen et al[14], 2016	HD-WLE	SVM	100 (60 early BE neoplasia/40 BE)	100 (60 early BE neoplasia/40 BE)	Sensitivity 83%/specificity 83%
de Groof et al[15], 2019	HD-WLE	SVM	60 (40 early BE neoplasia/20 BE)	60 (40 early BE neoplasia/20 BE)	Sensitivity 95%/specificity 85%
Ebigbo et al[16], 2019	HD-WLE	CNN	100 (50 early BE neoplasia/50 BE)	100 (50 early BE neoplasia/50 BE)	Sensitivity 92%/specificity 100%
Ebigbo et al[16], 2019	HD-WLE/NBI	CNN	148 (early BE neoplasia/BE)	148 (early BE neoplasia/BE)	HD-WLE sensitivity 97%/specificity 88%; NBI sensitivity 94%/specificity 80%
Hashimoto et al[17], 2020	WLE/NBI	CNN	1374 (early BE neoplasia/BE)	253 WLE (146 early BE neoplasia/107 BE) 205 NBI (79 early BE neoplasia/126 BE)	WLE sensitivity 98.6%/specificity 88.8%; NBI sensitivity 92.4%/specificity 99.2%
de Groof et al[18], 2020	HD-WLE	ResNet-UNet	1247 WLE + 297 HD-WLE (early BE neoplasia/BE)	80 (40 early BE neoplasia/40 BE)	Sensitivity 90%/specificity 88%
Swager et al[22], 2017	VLE	SVM	60 (30 early BE neoplasia/30 BE)	60 (30 early BE neoplasia/30 BE)	Sensitivity 90%/specificity 93%
Veronese et al[20], 2013	CLE	SVM	337 (23 GM/263 IM/51 neoplasia)	337 (23 GM/263 IM/51 neoplasia)	Sensitivity 96%/95%/100%
Ghatwary et al[23], 2019	CLE	SVM	262 (GM/IM/neoplasia)	262 (GM/IM/neoplasia)	Sensitivity 70%/93%/93%
Hong et al[24], 2017	CLE	CNN	236 (26 GM/155 IM/55 neoplasia)	26 (4 GM/17 IM/5 neoplasia)	Sensitivity 0%/100%/80%
Ebigbo et al[25], 2020	HD-WLE	CNN	129 (early BE neoplasia/BE)	62 (36 early BE neoplasia/26 BE)	Sensitivity 83.7%/specificity 100%
Cai et al[26], 2019	WLE	CNN	2428 (1332 early ESCC/1096 healthy control)	187 (91 early ESCC/96 healthy control)	Sensitivity 97.8%/specificity 85.4%
Ohmori et al[27], 2020	WLE/NBI	Single Shot MultiBox Detector	22562 (17435 superficial ESCC/5127 control)	727 (255 WLE/268 non-magnifying NBI/204 magnifying NBI)	WLE sensitivity 90%/specificity 76%; non-magnifying NBI sensitivity 100%/specificity 63%; magnifying NBI sensitivity 98%/specificity 56%
Zhao et al[31], 2019	Magnifying NBI	Double-labeling fully convolutional network	1383 (207 type A IPCL/970 type B1 IPCL/206 type B2 IPCL)	1383 (207 type A IPCL/970 type B1 IPCL/206 type B2 IPCL)	Sensitivity 71.5%/91.1%/83.0%
Nakagawa et al[32], 2019	WLE/NBI	CNN	8660 non-magnifying (7230 EP-SM1/1430 SM2/3); 5678 magnifying (4916 EP-SM1/762 SM2/3)	914 (405 non-magnifying/509 magnifying)	Non-magnifying sensitivity 95.4%/specificity 79.2%; magnifying sensitivity 91.6%/specificity 79.2%
Tokai et al[33], 2020	WLE/NBI	CNN	1751 superficial ESCC	291 (201 EP-SM1/90 SM2)	Sensitivity 84.1%/specificity 73.3%
Shin et al[36], 2015	HRME	Two-class linear discriminant analysis	104 (15 early ESCC/89 control)	167 (19 early ESCC/148 control)	Sensitivity 84%/specificity 95%
Quang et al[37], 2016	HRME	Two-class linear discriminant analysis	104 (15 early ESCC/89 control)	3 (1 early ESCC/2 control)	Sensitivity 100%/specificity 100%
Everson et al[38], 2019	magnifying NBI	CNN	7046 sequential images (squamous cell neoplasia/healthy control)	7046 sequential images (squamous cell neoplasia/healthy control)	Sensitivity 89.7%/specificity 96.9%
Guo et al[39], 2020	NBI	SegNet	6473 (early ESCC/control)	47 (27 non-magnifying videos/20 magnifying videos)	Non-magnifying NBI sensitivity 60.8%; magnifying NBI sensitivity 96.1%

AI: Artificial intelligence; EC: Esophageal cancer; HD-WLE: High-definition white light endoscopy; SVM: Support vector machine; BE: Barrett's esophagus; CNN: Convolutional neural network; NBI: Narrow band imaging; VLE: Volumetric laser endomicroscopy; CLE: Confocal laser endomicroscopy; GM: Gastric metaplasia; IM: Intestinal metaplasia; ESCC: Esophageal squamous cell carcinoma; IPCL: Intrapapillary capillary loop; EP-SM1: Epithelium-submucosal cancers invading up to 200 μm; HRME: High-resolution microendoscopy.