A User Interface for Automatic Polyp Detection Based on Deep Learning with Extended Vision

Colorectal cancer (CRC) is a leading cause of cancer-related deaths worldwide. To prevent CRC, the best method is screening colonoscopy. During this procedure, the examiner searches for colon polyps. Colon polyps are mucosal protrusions that protrude from the intestinal mucosa into the intestinal lumen. During the colonoscopy, all those polyps have to be found by the examiner. However, as the colon is folding and winding, polyps may hide behind folds or in uninvestigated areas and be missed by the examiner. Therefore, some publications suggest expanding the view of the examiner with multiple cameras. Nevertheless, expanding the examiner's view with multiple cameras leads to overwhelming and cumbersome interventions. Therefore, we suggest maintaining the examiner's classical endoscope view but extending the endoscope with side cameras. Those side camera views are only shown to an Artificial Intelligence (AI) trained for polyp detection. This AI system detects polyps on the side cameras and alarms the examiner if a polyp is found. Therefore, the examiner can easily move the main endoscope view on the AI detected area without being overwhelmed with too many camera images. In this study, we build a prototype of the endoscope with extended vision and test the automatic polyp detection system on gene-targeted pigs. Results show that our system outperforms current benchmarks and that the AI is able to find additional polyps that were not visualized with the main endoscope camera.