Aims The use of artificial intelligence (AI) may be an objective and operator-independent approach to increase endoscopist’s adenoma detection rate (ADR) and limit inter-operator variability. Within this study, we developed a deep neural network (DNN) for automated de-tection of colorectal polyps and assessed its performance for real-time detection and localiza-tion when incorporated into existing colonoscopy platforms.
Methods We collected a total of 116.529 colonoscopy images from 278 patients with a total of 788 polyps. Within this data set five expert endoscopists annotated the presence of polyps. These annotations were treated as the gold standard. A total of 10.467 annotated images from 504 different polyps were used as a training data set to generate the DNN. In order to assess the DNN’s performance a set of 45 videos comprising ~16.000 annotated video frames were used as a test set.
Results Half of the polyps in the test data were of flat morphology (50% Paris IIa/IIb) and either diminutive (≤5 mm, 52.5%) or small (6 to 10 mm, 47.5%) with an average size of 6 mm. Over 70% of the polyp annotations covered less than 5% of the entire video frame size further indicating that the analyzed polyps were not prominently visible within the video frame. The DNN’s sensitivity for polyp detection and localization was 90% with a specificity of 80% respectively (at 30Hz). In a receiver operating characteristic (ROC) analysis the system achieved an area under curve (AUC) of 92%.
Conclusions We generated a sensitive DNN for the automated detection and localization of colorectal polyps in real-time that holds the potential to be incorporated into existing colonoscopy platforms. In the near future a multi-center study will be conducted to further investigate the system’s effect on endoscopist’s ADR in vivo.
