Minireviews
Copyright ©The Author(s) 2021.
Artif Intell Gastrointest Endosc. Oct 28, 2021; 2(5): 198-210
Published online Oct 28, 2021. doi: 10.37126/aige.v2.i5.198
Table 1 Application of artificial intelligence in endoscopic detection of Barrett’s esophagus
Ref.
Target disease
Endoscopic modality
AI technology
Database
Outcomes
van der Sommen et al[23], 2016Early neoplasia in BEWLISVM100 imagesPer-image sensitivity 83%/specificity 83%; Per-patient sensitivity 86%/specificity: 87%
Struyvenberg et al[24], 2021BEWLI/NBICNNTrain 494364 images/1247 images; test 183 images/157 videosImages: Accuracy 84%/sensitivity 88%/specificity 78%; Videos: Accuracy 83%/sensitivity 85%/specificity 83%
de Groof et al[25], 2020Early neoplasia in BEWLIResNet-UNetTrain 1544 images; test 160 imagesDataset 4: Accuracy 89%/sensitivity 90%/specificity 88%; Dataset 5: Accuracy 88%/sensitivity 93%/specificity 83%
de Groof et al[26], 2020Barrett’s neoplasiaWLIResNet-UNetTrain 1544 images; test 20 patientsAccuracy 90%/sensitivity 91%/specificity 89%
Hong et al[27], 2017BEEndomicroscopyCNNTrain 236 images; test 26 imagesAccuracy 80.77%
Hashimoto et al[28], 2020Early neoplasia in BEWLI/NBICNNTrain 1832 images; test 458 imagesAccuracy 95.4%/sensitivity 96.4%/ specificity 94.2%
de Groof et al[29], 2019Barrett’s neoplasiaWLISVM60 imagesAccuracy 92%/sensitivity 95%/specificity 85%
Table 2 Application of artificial intelligence in endoscopic detection of early esophageal cancer
Ref.
Target disease
Endoscopic modality
AI technology
Database
Outcomes
Ebigbo et al[37], 2019EACWLI/NBICNN248 imagesAugsburg database: Sensitivity 97%/specificity 88% (WLI); Sensitivity 94%/specificity 80% (NBI); MICCAI database: Sensitivity 92%/specificity 100%
Ebigbo et al[38], 2020EACWLICNNTrain 129 images; test 62 imagesAccuracy 89.9%/sensitivity 83.7%/specificity 100%
Horie et al[39], 2019ECWLI/NBICNNTrain 8428 images; test 1118 imagesAccuracy 98%/sensitivity 98%
Cai et al[40], 2019ESCCWLIDNNTrain 2428 images; test 187 imagesAccuracy 91.4%/sensitivity 97.8%/specificity 85.4%
Ohmori et al[41], 2020ESCCWLI/NBI/BLICNNTrain 22562 images; test 727 imagesNon-ME: Accuracy 81.0%/sensitivity 90%/specificity 76% (WLI); Accuracy 77%/sensitivity 100%/specificity 63% (NBI/BLI); ME: Accuracy 77%/sensitivity 98%/specificity 56%
Liu et al[42], 2020ECWLICNNTrain 1017 images; test 255 imagesAccuracy 85.83%/sensitivity 94.23%/specificity 94.67%
Kumagai et al[43], 2019ESCCECSCNNTrain 4715 images; test 1520 imagesAccuracy 90.9%/sensitivity 92.6%/specificity 89.3%
Guo et al[44], 2020ESCCNBICNNTrain 6473 images; test 6671 images and 80 videosImages: Sensitivity 98.04%/specificity 95.03%; videos: Non-ME sensitivity 60.8% (per frame)/100% (per lesion); ME sensitivity 96.1% (per frame)/100% (per lesion)
Tokai et al[46], 2020ESCCWLI/NBICNNTrain 1751 images; test 291 imagesAccuracy 80.9%/sensitivity 84.1%/specificity 73.3%
Nakagawa et al[47], 2019ESCCWLI/NBICNNTrain 14338 images; test 914 imagesAccuracy 91%/sensitivity 90.1%/specificity 95.8%
Zhao et al[48], 2019ESCCNBIDouble-labeling FCN1350 imagesLesion level: Accuracy 89.2%; pixel level: Accuracy 93%
Everson et al[49]ESCCNBICNN7046 imagesAccuracy 93.7%/sensitivity 89.3%/specificity 98%
Uema et al[50], 2021ESCCNBICNNTrain 1777 images; test 747 imagesAccuracy 84.2%
Fukuda et al[51], 2020ESCCNBI/BLICNNTrain 28333 images; test 144 patientsAccuracy 63%/sensitivity 91%/specificities 51% (detection); accuracy 88%/sensitivity 86%/specificities 89% (characterization)
Shimamoto et al[52], 2020ESCCWLI/NBI/BLICNNTrain 23977 images; test 102 videosNon-ME: Accuracy 87%/sensitivity 50%/specificity 99%; ME: Accuracy 89%/sensitivity 71%/specificity 95%
Waki et al[53], 2021ESCCWLI/NBI/BLICNNTrain 18797 images; test 100 videosSensitivity 85.7%/specificity 40%