Fig 1.
Flowchart of the study design.
Fig 2.
Class activation map of inflamed mucosa.
Jet color map showing the normalized prediction where reddish and bluish colors are close to 1 and 0, respectively (right bar). An ulcer in the left, lower corner of the image is highlighted in red. This figure demonstrates the approximate mechanism of AI.
Table 1.
Hyperparameters for the training of the AI.
Fig 3.
Concordance of significant images between the AI and experts.
All the four images were classified as significant with 0.8 or higher probability and based on manual classification by experts (left side of the images). As the color spectrum of class activation map turns red, lesions display higher probabilities. AI can distinguish multiple findings that coexist in an image (right side of the images; A, swollen villi from debris; B, Small mucosal defect from the nearby debris; C, Vascular tuft adjacent to vessels; D, Vascular tuft surrounded by inflamed mucosa).
Fig 4.
Receiver operating characteristic curve of the AI for binary classification.
The receiver operating characteristic (ROC) curve of AI for detection of significant images: Area under the curves (AUCs) were 0.9982, 0.9981, and 0.9999 for test, validation, and training set images, respectively. ROC curve and AUCs shown that the training model is well fitted to all of training images as well as there is little degradation of validation and testing performance from the training model.
Table 2.
Clinical characteristics of cases and summary result of AI.
Table 3.
Comparison of the lesion detection rates and reading times of reviewers between the reading models.
Table 4.
Comparison of the lesion detection rates and reading times of reviewer groups between the reading models.
Fig 5.
Improvement in the capsule endoscopy reading using the AI-assisted reading model.