| Mi et al., 2020 [47] | South Korea | 1169 patients | Narrow-band imaging (NBI) images of colorectal polyps | Computer-aided diagnosis system (CAD) | CAD + trained examiner achieved a Cohen’s kappa value of 0.665 (95% CI, 0.560 - 0.758), higher than the examiner alone (0.368, 95% CI 0.281 - 0.459). CAD + trained examiner had a significantly higher overall diagnostic accuracy compared to trainee endoscopists (84.2% vs. 71.8%; p < 0.001). | AI CAD system can assist inexperienced endoscopists in accurately predicting the histopathology of colorectal polyps with over 80% diagnostic accuracy, regardless of size, location, or morphology. |
Wang et al., 2020 [49] | China | 240 patients | High-resolution axial T2WI MRI images | Faster R-CNN model | The accuracy, sensitivity, and specificity of the model in identifying positive circumferential resection margin (CRM) status in high-resolution MRI images were 0.932, 0.838, and 0.956, respectively. The proportion of positive CRM images with an overlap ratio above 0.7 with the region annotated by radiologists was 95.1% (175/184). | The trained Faster R-CNN AI method’s ability to annotate CRM was comparable to radiology experts and had higher efficiency in identification. | |
Lu et al., 2018 [50] | China | 28,080 images | MRI images of metastatic lymph nodes | Faster R-CNN | The area under the receiver operating characteristic (ROC) curve for Faster R-CNN was 0.912. | The ability of Faster R-CNN to label lymph node positions was similar to that of senior radiologists. |