Table 1.
Single-frame algorithms for anatomical landmark detection.
| Authors | Data Annotation Protocol | Dataset | Classes | Algorithm | Pre-Processing | Validation | Average Performance | AI Impact (Clinical Setting) |
|---|---|---|---|---|---|---|---|---|
| Takiyama et al., 2018 [39] | Undefined | (Private Dataset) 44,416 UGIE images (optimal view WLI) | 4 sites + 3 gastric sites | GoogleLeNet | Black frame cropping | Holdout set | (4 anatomical classes) Accuracy: 99% Sensitivity: 94% (from 87% to 99%) 96% (from 96% to 97%) (3 gastric sub-classes) Sensitivity: 97% (from 96% to 97%) Specificity: 98% (from 98% to 99%) |
N/A |
| Wu et al., 2019 [40] | 2 experts with >10 years of experience | (Private Dataset) 24,549 WLI images |
10 or 26 sites | VGG16-Resnet50 | CNN filters blurry frames | Holdout set | Accuracy: 90% (10 sites) 66% (26 sites) |
(Single-center, retrospective trial) [40] **** Endoscopist accuracy: 90% (10 sites, experts) 63% (26 sites, experts) 87% (10 sites, seniors) 59% (26 sites, seniors) 83% (10 sites, novices) 46% (26 sites, novices) |
| Zhang Xu et al., 2019 [41] | 2 expert endoscopists (years of experience unknown) | (Private Dataset) 75,275 UGIE images (including non-informative and NBI frames) * |
10 sites + uninformative + NBI | Muli-Task Custom CNN + SSD | None | Holdout set | Average precision (mAP): 94% |
N/A |
| He et al., 2019 ** [42] | 1 doctoral student 1 clinical gastroenterology research fellow |
3704 UGIE images (WLI+LCI frames) optimal views | 11 sites + N/A | Inception-v3 | Data-driven ROI cropping | 5-fold C.V. | Accuracy: 83% F1: 80% (from 53% to 94%) |
N/A |
| Igarashi et al., 2020 [43] | 1 expert with >30 years of experience 1 endoscopists with >4 years of experience |
(Private Dataset) 85,246 upper GI images |
10 sites from UGIE + 4 classes pertaining to specimens and other examinations | AlexNet | None | Holdout set | Accuracy: 97% |
N/A |
| Sun et al., 2021 *** [44] | >1 endoscopist with >5 years of experience | (Private Dataset) 10,474 UGIE images including NBI |
11 sites + NBI | Custom CNN+RCF | ROI extraction + bilinear interpolation | 5-fold C.V. | Accuracy: 99% Precision: 93% F1 score: 92% |
N/A |
| Chang et al., 2021 [27] | Unclear | (Private Dataset) 15,723 frames from asymptomatic patients |
8 classes | ResNeSt | None | Holdout set | Accuracy: 97% |
N/A |
* Addressed in different classification tasks. ** We only consider the protocol with all the landmarks and N/A classes in this article. *** We considered only the values for the first ResNeSt, since the ampulla was divided into two categories and trained with a second model and different dataset. **** Only findings concerning blind spots were considered.