Table 4.
Studies on the use of Artificial Intelligence (AI) in conjunction with Endoscopic Ultrasound (EUS) for the detection of Gastrointestinal Stromal Tumors (GISTs).
| Author | Study | N. EUS Images GISTs | N. GISTs | AI System | Lesion Size mm | Sensitivity | Specificity | AUROC | Conclusion |
|---|---|---|---|---|---|---|---|---|---|
| Kim YH et al., 2020 [102] | Retrospective | 905 images of gastric mesenchymal tumors (GIST, leiomyoma, and schwannoma): training dataset; 212 images of gastric mesenchymal tumors: valdation | Training dataset: 125 (69.8%); test dataset: 32 (46.4%) | CNN-CAD system | Training dataset: 3.6 ± 2.1; Test dataset: 3.2 ± 1.6 | 83.0 (77.4–87.5) | 75.5 (69.3–80.8) | 79.2 (73.3–84.2) | The CNN-CAD system performed exceptionally well with respect to detecting gastric mesenchymal tumors. |
| Oh CK et al., 2021 [103] | Retrospective | 376 images (n = 114 pts) | Training dataset: 85; validation dataset: 54 | CNN-based object | 25 (10–70) | 100% (per-patient) | 85.7% (per-patient) | 96.3% (per-patient) | High diagnostic ability for predicting gastric GISTs and outperformed human assessment. |
| Hirai K et al., 2022 [104] | Retrospective | 16,110 images (n = 631 pts) | Training dataset: 287 (68.5); validation dataset: 63 (70.0); test dataset: 85 (69.7) | AI—deep learning | Training: 25 (2.2–180); validation: 28 (6–130); test: 26.1 (3–180) | 98.8% | 67.6% | 89.3% | In terms of diagnostic performance, the AI system that classified SELs outperformed the experts and may help improve SEL diagnosis in clinical practice. |
| Yang X et al., 2022 [105] | Retrospective | 10,439 images (n = 752 pts) | 36 | AI-based system | Endosonographers’ accuracy in diagnosing GISTs or GI leiomyomas increased from 73.8% (95%CI 63.1–82.2%) to 88.8% (95%CI 79.8–94.2%; p = 0.01) | An AI-based EUS diagnostic system was developed that can effectively distinguish GISTs from GI leiomyomas and improve the diagnostic accuracy of SEL assessment. | |||
| Tanaka H et al., 2022 [106] | Retrospective | 10,600 images (n = 53 pts) | 42 | AI—deep learning involving a residual neural network and leave-one-out cross-validation | 26.4 | The sensitivity, specificity, and accuracy of AI for diagnosing GISTs were 90.5%, 90.9%, and 90.6%, which can be compared to 90.5%, 81.8%, and 88.7%, respectively, obtained for blind reading (p = 0.683) | The diagnostic ability of AI-evaluated CH-EUS results to distinguish between GISTs and leiomyomas was comparable to blind reading by expert endosonographers. | ||
| Liu XY et al., 2022 [107] | Meta-analysis (8 studies) | NA | 339 (training, validation, and test datasets) | Convolutional neural network (CNN) model | In terms of sensitivity (0.93 vs. 0.71), specificity (0.81 vs. 0.69), and AUC (0.94 vs. 0.75), AI-aided EUS outperformed expert-conducted EUS | AI-assisted EUS is a promising and dependable method for separating SELs with excellent diagnostic performance | |||
Abbreviations: GISTs: Gastrointestinal stromal tumors; SELs: Subepithelial lesions; GI: gastrointestinal; CNN-CAD: Convolutional neural network computer-aided diagnosis.