IMPROVING USABILITY OF AI SYSTEMS FOR POLYP DETECTION BY RECOGNIZING DIFFERENT INTERVENTIONS DURING COLONOSCOPY

Aims Artificial intelligence for polyp detection systems (CADe) highlight regions such as colorectal polyps and are useful in fully insufflated colon lumen. However, CADe generate a significant number of false positive (FP) activations when performing interventions such as polypectomies due to the introduction of snare or biopsy forceps. These bounding boxes have the potential to disturb the examiner’s work.

Methods A convolutional neuronal network (CNN) to recognize instruments in the endoscopic image was developed and evaluated. The CNN has the ability to pause the signal of the CADe system when an instrument is recognized. A total of 30 different examinations from 6 different centers were screened for instruments and generated the training dataset. The test dataset included 8 full-colonoscopy videos that were analyzed for the recognition of visible instruments and detections by a commercially available CADe.

Results The training data contained 74179 images, 23.9% with visible instruments. The CNN was able to recognize instruments in 73.5% of the validation dataset images with a specificity of 90.9%. A mean of 380.5 disturbing frames per colonoscopy were avoided using the CNN. This accounted for a 76.9% of the total number of disturbing activations.

Conclusions CADe systems usually rely on a clean, well-insufflated colon lumen to detect polyps. However, instruments like polypectomy snares often lead to FP detections that could potentially disturb the examiner during the intervention. Using a CNN, we were able to accurately detect the presence of an instrument, pause the CADe system and avoid further activations when the polyp is already detected.

Publication History

Article published online:
14 April 2022