Table 2.
Key studies on AI application for capsule endoscopy (CE) in IBD.
| Author (Year) | Study design | Population | Outcome | Results |
|---|---|---|---|---|
| Aoki et al. (36) | Retrospective cohort study | 10,440 CE images | To assess a CNN system for automated identification of ulcers and erosions in CE images of SB | The CNN evaluated 10,440 images in 233 seconds and identified ulcers and erosions with 88.2% sensitivity and 90.9% specificity |
| Aoki et al. (37) | Retrospective cohort study | 20 entire SB CE videos | To evaluate a CNN model as the first screening on SB CE video readings, comparing endoscopist reviewing after the CNN screening with endoscopist-alone reviewing | CNN reduced reviewing time (from 12.2 min to 3.1 for experienced operators and from 20.7 to 5.2 for trainees) without affecting detection rate of erosions and ulcers (experienced operators: 87 vs. 84%; trainees: 55 vs. 47%) |
| Klang et al. (38) | Retrospective cohort study | 17,640 CE images from 49 CD patients | To test a CNN system for the automated identification of SB ulcers in CD on CE images | The CNN algorithm discriminated normal mucosa from ulcers with high accuracy (>95%) |
| Barash et al. (39) | Retrospective cohort study | 17,640 CE images from 49 CD patients | To assess a CNN for grading CD ulcers on CE images | The AI-assisted tool had an overall agreement with capsule readers of 67%, with an accuracy of 91% for severe ulcers |
| Ferreira et al. (40) | Retrospective cohort study | 8,085 CE images from CD patients | To evaluate an AI algorithm for the automated detection of erosions and ulcerations in both SB and colon CE images from CD patients. | The CNN system accurately identified both ulcers (sensitivity 83%; specificity 98%) and erosions (sensitivity 91%; specificity 93%) |
| Xie et al. (41) | Retrospective cohort study | 2,898 CE videos | CAD system trained on CE videos vs. conventional reading, in detection and classification of SB findings | The DNN-based reading reached higher detection rate of SB findings than conventional reading (95.9 vs. 76.1%) in a less time (5.4 vs. 51.4 min) |
AI, artificial intelligence; CAD, computer-assisted diagnosis; CD, Crohn's disease; CE, capsule endoscopy; CNN, convolution neural network; DNN, deep neural network; SB, small bowel.