| Mascarenhas et al., 2021 [43] | Portugal | 53,555 images | CE images | CNN Model | The average sensitivities and specificities for automatically detecting various abnormalities were 87.8% ± 8.1% and 99.4% ± 3.7%, respectively. | Deep learning algorithms can detect and differentiate small bowel lesions with obvious bleeding potential. |
Ungaro et al., 2020 [44] | United States | 265 patients with an average age of 11.6 years | Protein and serologic markers | Multivariate Cox regression model | For B2 (stricturing) complications, four proteins (IL7, MMP10, IL12B, CCL11) and two serologic markers (LnASCA IgA, LnCbir) were selected as the most predictive. For B3 (penetrating) complications, five proteins (TNFSF14, CCL4, IL15RA, TNFB, CD40) and three serologic markers (LnASCA IgA, LnANCA, LnCbir) were selected as the most predictive for B3. | Different proteins have better predictive value for diagnosing different complications of pediatric Crohn’s disease, B2 and B3. | |
Colorectal Cancer | Kudo et al., 2020 [46] | Japan | 69,142 images | WLI images, EC-NBI images, and stained EC images | EndoBRAIN CAD system | In stained images: EndoBRAIN achieved an accuracy of 98% (97.3 - 98.6). In NBI images: EndoBRAIN achieved an accuracy of 94.6%. | EndoBRAIN improved the accuracy and sensitivity of endoscopists in both stained and NBI modes. |