| Shahriar et al., 2022 [22] | USA | 542 cases | Whole slide imaging | DL | ResNet101 model: Sensitivity and specificity for LGD are 81.3% and 100% respectively, while for NDBE and HGD they are both > 90%. | ResNet101 model can predict dysplasia grade on whole slide imaging. |
Wen et al., 2022 [23] | China | 187 images | Endoscopic images | Fully Convolutional Networks (FCN) | Intersection over Union (IOU) values of 0.56 (GEJ) and 0.82 (SCJ). | The segmentation results of fully automatic DL method are consistent with manual evaluation. | |
Alanna et al., 2020 [24] | Germany | 129 images | Endoscopic images | CNN and DeepLab V.3+ residual network (ResNet) architecture | AI system has sensitivity and specificity of 83.7% and 100.0% respectively, with an overall accuracy of 89.9%. | This is the first real-time application of deep learning AI system for evaluating and diagnosing early EAC in real-life scenarios. | |
Manon et al., 2021 [25] | Netherlands | 57 cases | Mass spectrometry imaging (MSI) and hematoxylin and eosin (H & E) staining imaging | ML classifier | Differentiating epithelial tissue from stroma: AUC of 0.89 (MSI) and 0.95 (H & E); distinguishing dysplasia grade: AUC of 0.97 (MSI) and 0.85 (H & E); low-grade progressors and non-progressors: accuracy of 0.72 (MSI) and 0.48 (H & E). | H & E-based classifier excels in differentiating tissue types, while MSI-based model is more accurate in distinguishing dysplasia grade and risk of progression. | |
Sharib et al., 2021 [26] | UK | 131 videos | Endoscopic high-definition videos | Depth Estimation Networks | Accuracy of phantom endoscopic videos for C & M and island measurements is 97.2%, with an average deviation of ±0.9 mm, while for BEA it is 98.4% with an average deviation of ±0.4 cm. | The quantification system can automatically measure C & M score, quantify Barrett’s epithelium area (BEA), and measure island area, enabling esophageal 3D reconstruction. |