| Chierici et al., 2022 [40] | Italy | 14,226 images | Endoscopy images | Residual Network (ResNet) | The combination model resnet 34-50-101, resnet 34-50-152 exhibited more prominent recognition capabilities in N-P, UC-CD, and UC-N samples compared to any individual resnet model. | Ensemble learning methods effectively improve the performance of the combination model. |
Eyal et al., 2020 [41] | Israel | 27,892 images | CE images | Deep Learning Network | The network achieved an average accuracy of 93.5% in classifying narrow and non-narrow cases. It also demonstrated good discrimination between narrow and normal mucosa, narrow and all ulcers, as well as narrow and different grades of ulcers. | The DL network was validated to effectively identify intestinal strictures, normal mucosa, and different grades of ulcers in Crohn’s disease patients’ CE images. | |
Barash et al., 2021 [42] | Israel | 17,640 images | CE images | CNN | The accuracy for classifying grade 1 and grade 3 ulcers was 0.91 (95% confidence interval, 0.867 - 0.954). For grade 2 and grade 3, it was 0.78 (95% confidence interval, 0.716 - 0.844), and for grade 1 and grade 2, it was 0.624 (95% confidence interval, 0.547 - 0.701). | The CNN achieved a high accuracy in detecting severe CD ulcers but performed poorly in mild CD ulcers. |