Emerging role of artificial intelligence in gastroenterology and hepatology

Table 1 Use of artificial intelligence in the identification of patients with esophageal neoplastic or preneoplastic lesions

Lesions	Diagnostic or predictive modality	AI classifier	AI validation methods	Number of images, slides or videos in training dataset	Number of images, slides or videos in test dataset	Best average results (%)		Ref.
						Accuracy	Sensitivity/specificity
NDBE, LGD, HGD	Histology	Deep learning	Training, validation, and test sets created (by a random 70/20/10 split) from 542 patients (164 NDBE, 226 LGD, and 152 HGD)	8596 bounding boxes	840 boxes	F1 score; NDBE: 0.91; LGD: 0.90; HGD: 1.0	NDBE: > 90; LGD: 81.3/100; HGD: > 90	Faghani et al[18]
ESCC and EAC	Upper GI endoscopy (WL or NBI)	CNN		8428 from 384 patients (397 ESCC, 32 EAC)	1118 (956/162) from 50 control, 41 ESCC, 8 EAC	55.7	98.0/16.0	Horie et al[19]
EN-BE	Upper GI endoscopy	Deep learning CAD system		494364 Labeled endoscopic images collected from all intestinal segments	1704 unique esophageal high-resolution images of rigorously confirmed EN-BE and NDBE, derived from 669 patients	89.0	88.0/90.0	de Groof et al[20]
EN-BE	Upper GI endoscopy	CNN-CAD	Phase 2 image-based validation; phase 3 video-based external validation	75198 images and videos (96 patients) of neoplastic and 1014973 images and videos (65 patients) of nonneoplastic BE	Phase 2 107 images (20 patients) of neoplastic and 364 images (14 patients) of nonneoplastic BE; phase 3 32 videos (32 patients) of neoplastic and 43 videos (43 patients) of nonneoplastic BE	Phase 2: 94.7	95.3/94.5	Abdelrahim et al[23]
						Phase 3: 92.0	93.8/90.7

AI: Artificial intelligence; NDBE: Non-dysplastic Barrett’s esophagus; LGD: Low grade dysplasia; HGD: High grade dysplasia; ESCC: Esophageal squamous cell carcinoma; EAC: Esophageal adenocarcinoma; EN-BE: Early-stage neoplasia in patients with Barrett’s esophagus; GI: Gastrointestinal; NBI: Narrow-band imaging, WL: White light; CNN: Convolutional neural network; CAD: Computer-aided detection; BE: Barrett’s esophagus.

Table 2 Use of artificial intelligence in the identification of patients with early gastric cancer

Lesions	Diagnostic or predictive modality	AI classifier	Number of images in training dataset	Number of images in test dataset	Best average results (%)		Ref.
					Accuracy	Sensitivity/specificity
EGC	Upper GI endoscopy (WL, CE, NBI)	CNN	13584 from 2639 lesions	2296 from 77 lesions	NA	92.2/NA	Hirasawa et al[27]
EGC	Upper GI endoscopy	CNN-CAD system	790 images	203 images	89.16	76.47/95.56	Zhu et al[28]
EGC	Histology	CNN	2123 whole slide images	3212 whole slide images		100/80.6	Song et al[29]

AI: Artificial intelligence; EGC: Early-stage gastric cancer; GI: Gastrointestinal; NBI: Narrow-band imaging, WL: White light; CE: Chromoendoscopy; CNN: Convolutional neural network; CAD: Computer-aided detection; NA: Not available.