Deep learning for diagnosis of precancerous lesions in upper gastrointestinal endoscopy: A review

Table 1 Summary of studies using deep learning for detection of esophageal precancerous lesions

Ref.	Year	Imaging	Study design	Study aim	DL model	Dataset	Outcomes
Cai et al[30]	2019	WLE	Retrospective	Detection of precancerous lesions and early ESCC	--	2615 images	Sensitivity: 97.8%. Specificity: 85.4%. Accuracy: 91.4%
Guo et al[31]	2020	NBI, M-NBI	Retrospective	Detection of precancerous lesions and early ESCC	SegNet	13144 images and 168865 video frames	Sensitivity: 96.10% for M-NBI videos, 60.80% for non-M-NBI videos, 98.04% for images. Specificity: 99.90% for non-M-NBI/M-NBI videos, 95.30% for images
de Groof et al[32]	2020	WLE	Retrospective	Detection of Barrett’s neoplasia	ResNet/U-Ne	1544 images	Sensitivity: 91%. Specificity: 89%. Accuracy: 90%
de Groof et al[33]	2020	WLE	Retrospective	Detection of Barrett’s neoplasia	ResNet/U-Ne	494364 unlabeled images and 1704 labeled images	Sensitivity: 90%. Specificity: 88%. Accuracy: 89%
Struyvenberg et al[34]	2021	NBI	Retrospective	Detection of Barrett’s neoplasia	ResNet/U-Ne	2677 images	Sensitivity: 88%. Specificity: 78%. Accuracy: 84%
Hashimoto et al[35]	2020	WLE, NBI	Retrospective	Recognition of early neoplasia in BE	Inception-ResNet-v2, YOLO-v2	2290 images	Sensitivity: 96.4%. Specificity: 94.2%. Accuracy: 95.4%
Hussein et al[36]	2020	WLE	Retrospective	Diagnosis of early neoplasia in BE	Resnet101	266930 video frames	Sensitivity: 88.26%. Specificity: 80.13%
Ebigbo et al[37]	2020	WLE	Retrospective	Diagnosis of early EAC in BE	DeepLab V.3+, Resnet101	191 images	Sensitivity: 83.7%. Specificity: 100%. Accuracy: 89.9%
Liu et al[38]	2020	WLE	Retrospective	Detection of esophageal cancer from precancerous lesions	Inception-ResNet	1272 images	Sensitivity: 94.23%. Specificity: 94.67%. Accuracy: 85.83%
Wu et al[39]	2021	WLE	Retrospective	Automatic classification and segmentation for esophageal lesions	ELNet	1051 images	Classification sensitivity: 90.34%. Classification specificity: 97.18%. Classification accuracy: 96.28%. Segmentation sensitivity: 80.18%. Segmentation Specificity: 96.55%, Segmentation accuracy: 94.62%
Ghatwary et al[40]	2021	WLE	Retrospective	Detection of esophageal abnormalities from endoscopic videos	DenseConvLstm, Faster R-CNN	42425 video frames	Sensitivity: 93.7%. F-measure: 93.2%

BE: Barrett’s esophagus; DL: Deep learning; EAC: Esophageal adenocarcinoma; ESCC: Esophageal squamous cell carcinoma; M-NBI: Magnifying narrow band imaging; NBI: Narrow band imaging; WLE: White light endoscopy.

Table 2 Summary of studies using deep learning for detection of gastric precancerous lesions

Ref.	Year	Imaging	Study design	Study aim	DL model	Dataset	Outcomes
Shichijo et al[47]	2017	WLE	Retrospective	Diagnosis of H. pylori infection	GoogLeNet	43689 images	Sensitivity: 88.9%; Specificity: 87.4%; Accuracy: 87.7%
Itoh et al[48]	2018	WLE	Retrospective	Analysis of H. pylori infection	GoogLeNet	179 images	Sensitivity: 86.7%; Specificity: 86.7%
Zheng et al[49]	2019	WLE	Retrospective	Evaluation of H. pylori infection status	ResNet-50	15484 images	Sensitivity: 91.6%; Specificity: 98.6%; Accuracy: 93.8%
Nakashima et al[50]	2018	BLI-bright, LCI	Prospective	Prediction of H. pylori infection status	GoogLeNet	666 images	Sensitivity: 96.7%; Specificity: 86.7%
Nakashima et al[51]	2020	WLE, LCI	Prospective	Diagnosis of H. pylori infection	--	13127 images	For currently infected patients, the sensitivity and specificity are 62.5% and 92.5%, respectively
Guimarães et al[53]	2020	WLE	Retrospective	Diagnosis of atrophic gastritis	VGG16	270 images	Accuracy: 93%
Zhang et al[54]	2020	WLE	Retrospective	Diagnosis of atrophic gastritis	DenseNet121	5470 images	Sensitivity: 94.5%; Specificity: 94.0%; Accuracy: 94.2%
Horiuchi et al[55]	2020	M-NBI	Retrospective	Differentiation between early gastric cancer and gastritis	GoogLeNet	2826 images	Sensitivity: 95.4%; Specificity: 71.0%; Accuracy: 85.3%
Wang et al[57]	2019	WLE	Retrospective	Localization and identification of GIM	DeepLab V.3+	200 images	Accuracy: 89.51%
Zheng et al[58]	2020	WLE	Retrospective	Detection of atrophic gastritis and GIM	ResNet-50	3759 images	Sensitivity for atrophic gastritis: 87.2%; Specificity for atrophic gastritis: 91.1%; Sensitivity for GIM: 90.3%; Specificity for GIM: 93.7%
Yan et al[18]	2020	NBI, M-NBI	Retrospective	Diagnosis of GIM	EfficientNetB4	2357 images	Sensitivity: 91.9%; Specificity: 86.0%; Accuracy: 88.8%
Cho et al[60]	2019	WLE	Prospective	Classification of multiclass gastric neoplasms	Inception-Resnet-v2	5217 images	Accuracy: 84.6%
Inoue et al[61]	2020	WLE, NBI	Retrospective	Detection of duodenal adenomas and high-grade dysplasias	Single-Shot Multibox Detector	1511 images	For high-grade dysplasia, the sensitivity and specificity are all 100%
Lui et al[62]	2020	NBI	Retrospective	Classification of gastric lesions	ResNet	3000 images	Sensitivity: 97.1%; Specificity: 85.9%; Accuracy: 91.0%

BLI-bright: Blue laser imaging-bright; DL: Deep learning; GIM: Gastric intestinal metaplasia; H. pylori: Helicobacter pylori; LCI: Linked color imaging; M-NBI: Magnifying narrow band imaging; NBI: Narrow band imaging; WLE: White light endoscopy.