ARTIFICIAL INTELLIGENCE COMBINED WITH LCI YIELDS IN HIGHEST ACCURACY AND DETECTION OF COLORECTAL POLYPS, INCLUDING SESSILE SERRATED LESIONS

Aims Linked color imaging (LCI) has shown its effectiveness in multiple randomized controlled trials for enhanced colorectal polyp detection. Most recently, artificial intelligence (AI) with deep learning through convolutional neural networks has dramatically improved and is increasingly recognized as a promising new technique enhancing colorectal polyp detection. Study aim was to evaluate a new developed deep-learning computer-aided detection (CAD) system in combination with LCI for colorectal polyp detection.

Methods First, a convolutional neural network was trained for colorectal polyp detection in combination with the LCI-technique using a dataset of anonymized endoscopy videos. For the validation, 240 polyps within fully recorded endoscopy videos with LCI mode, covering the whole spectrum of adenomatous histology, were used. Sensitivity (True positive rate per-lesion) and false positive frames in a full procedure were assessed.

Results The new CAD system used on LCI mode could at least process 60 frames per second allowing for real-time video analysis. Sensitivity (True positive rate per-lesion) was 100% with no lesion being missed. The calculated false positive frame rate was 0.001%. Out of the 240 polyps included, 34 were sessile serrated lesions. The detection rate for sessile serrated lesions with the CAD system used on LCI mode was 100%.

Conclusions The new CAD system used on LCI mode achieved a 100% sensitivity per lesion and a negligible false positive frame rate. Of note, the new CAD system used on LCI mode also specifically allowed for detection of serrated lesions in all cases.