. 2020 Oct 22;8(11):E1584–E1594. doi: 10.1055/a-1236-3007

Table 1. Study characteristics.

Study	Aim	Endoscopy technique	Machine learning model	Training strategy	Testing strategy	Accuracy	Sensitivity	Specificity	PPV	NPV	Remarks
Ahmad, 2019 ¹¹	Colorectal adenoma detection	Standard colonoscopy	CNN	Multicenter colonoscopy images and videos of 4664 polyp test frames	17 video datasets of complete colonoscopy withdrawal with 83 polyps consisting of 83716 frames (14634 polyp & 69082 non-polyp)	92.5	84.5	92.5	NR	NR	Conference abstract
Byrne, 2019 ²	Colorectal polyp detection in real-time endoscopic video images	NBI endoscopy	CNN	Unaltered video frames	125 videos of consecutively encountered diminutive polyps	94	98	83	90	97	–
Cai, 2019 ¹²	Detect early ESCC under conventional endoscopic white light imaging	White light endoscopy	CNN	1332 abnormal and 1096 normal images	187 images from 57 patients	91.4	97.8	85.4	86.4	97.6	–
Chen, 2018 ¹³	Colorectal polyp detection	NBI endoscopy	CNN (TensorFlow algorithm)	1476 images of neoplastic and 681 of hyperplastic polyps	96 hyperplastic and 188 neoplastic polyps smaller than 5 mm	90.1	96.3	78.1	89.6	91.5	–
Cho, 2019 ¹⁴	Classify gastric neoplasms based on endoscopic white-light images	White light endoscopy	CNN (Inception-v4, Resnet-152, Inception-Resnet-v2)	5017 images from 1269 individuals	200 images from 200 patients	93	60.7	98.3	85	93.9	Advanced gastric cancer
						74.5	28.3	88.3	41.9	80.5	Early gastric cancer
						86.4	0	99.4	0	86.9	High-grade dysplasia
						78.5	6.7	91.2	11.8	84.7	Low-grade dysplasia
						66.5	95.7	50.8	51.1	95.7	Non-neoplasm
Guo, 2019 ¹⁵	Real time automated diagnosis of precancerous and early ESCC in both non-magnifying and magnifying settings	NBI endoscopy	CNN (SegNet architecture)	2770 images of precancerous lesions and early ESCC in 191 cases and 3703 images of non-cancerous lesions in 358 cases	1480 malignant images in 59 cases, 5191 non-cancerous images in 2004 cases, 27 precancerous and early ESCC videos, and 33 normal videos	nr	98.04	95.03	NR	NR	–
Guo, 2019 ¹⁵		NBI endoscopy videos	CNN (SegNet architecture)			nr	60.8	99.9	NR	NR	–
Hirasawa, 2018 ¹⁶	Detect early and advanced gastric cancer	Standard white-light, chromoendoscopy, NBI	CNN (Single Shot MultiBox Detector)	13584 EGD images for 2639 histologically proven gastric cancer	2296 images from 77 gastric cancer lesions of 69 patients	nr	92.2	nr	30.6	NR	–
Horie, 2018 ¹⁷	Detect esophageal cancer	White-light, NBI	CNN (Single Shot MultiBox Detector)	8428 histologically proven EGD images of cancer in 384 patients	162 images of cancer and 376 images without cancer from 47 patients with 49 cancer lesions. 573 images of non-cancerous areas from 50 patients with no cancer	nr	77	79	39	95	–
Horiuchi, 2019 ¹⁸	Differentiate gastric cancer from gastritis	Magnifying NBI endoscopy	CNN (GoogLeNet)	1492 cancer and 1078 gastritis images	151 cancer and 107 gastritis images	85.3	95.4	71	82.3	91.7	–
Ikenoyama, 2019 ¹⁹	Detect gastric cancer	Standard EGD	CNN	13584 images from 2639 lesions	2940 images from 140 cases (209 early cancer images, 2731 non-neoplastic images)	nr	65.6	NR	14.6	NR	Conference abstract
Ito, 2018 ²⁰	Assist in cT1b colorectal cancer diagnosis	White-light colonoscopy	CNN (AlexNet & Caffe)	Group 1: 2520 cTis + cT1a, 2418 cT1b images; Group 2: 2604 cTis + cT1a, 2400 cT1b images; Group 3: 2604 cTis + cT1a, 2418 cT1b images	190 conventional white-light images	81.2	67.5	89	NR	NR	–
Komeda, 2019 ²¹	Colorectal polyp classification	White-light, NBI, chromo-endoscopy	CNN	29572 adenoma images, 62999 non-adenoma images	60 cases of colon polyps	nr	97.5	97.9	NR	NR	Conference abstract, white light
						nr	94.8	96.5	NR	NR	Conference abstract, NBI
						nr	90.1	99.5	NR	NR	Conference abstract, chromo-endoscopy
Li, 2019 ²²	Early gastric cancer detection	Magnifying NBI endoscopy	CNN (InceptionV3-Keras framework)	386 images of non-cancer lesions, 1702 images of early cancer	341 images (170 early cancer & 171 non-cancer lesions)	90.91	91.18	90.64	90.64	91.18	–
Liu, 2018 ²³	Early gastric cancer detection	Magnifying NBI endoscopy	CNN (VGG16, InceptionV3, InceptionResNetV2)	Magnifying NBI of normal gastric images and early gastric cancer images	Images (number not mentioned)	98.5	98.1	98.9	NR	NR	Conference abstract
Ozawa, 2018 ²⁴	Automatic endoscopic detection and classification of colorectal polyps	White-light & NBI colonoscopy	CNN (Single Shot MultiBox Detector)	16418 images of 4752 histologically proven colorectal polyps and 4013 images of normal colorectum	3533 images	nr	92	nr	93	NR	Conference abstract
Sakai, 2018 ²⁵	Detect early gastric cancer	White light endoscopy	CNN (Single Shot MultiBox Detector)	1000 images of 0-I, 0-IIa, 0-IIc lesions	228 cancer images	82.8	73.6	98.8	NR	NR	–
Wu, 2019 ⁵	Detect early gastric cancer	EGD	CNN (VGG-16, ResNet-50, Google's TensorFlow)	2204 early cancer, 326 advanced cancer, 4791 control	170 images	92.5	94	91	91.26	93.81	–
Zhang, 2017 ²⁶	Early esophageal neoplasia	NBI and magnifying endoscopy	CNN	218 endoscopic images generated 218000 patches, 90 % used for training and rest for testing		79.38	73.41	83.54	72.09	84.44	Conference abstract
Zhao, 2019 ⁶	Early ESCC using magnifying NBI	Magnifying NBI endoscopy	VGG16	three test groups with 463, 438 & 449 images	1350 images in 219 patients	89.2	87	84.1	81.9	90.4	–

CNN, convolutional neural networks; PPV, positive predictive value; NPV, negative predictive value; ESCC, esophageal squamous cell carcinoma; NBI, narrow band imaging; EGD, esophagogastroduodenoscopy; PET, positron emission tomography; ESD, endoscopic submucosal dissection; VGG, visual geometry group; NR, not reported