MISS RATE DUE TO FAILURE RECOGNITION: ARTIFICIAL INTELLIGENCE ESTIMATE BASED ON A HIGH-QUALITY. A RANDOMIZED CLINICAL TRIAL

Aims One fourth of polyps seems to be missed by“tandem”or“back-to-back”colonoscopy studies. However, this approach is not as optimal as prone to observer and procedural variability during the two consecutive procedures.Furthermore,it is unclear how much of this miss rate is actually due to failed recognition. Our aim is to use Artificial-Intelligence(AI)for estimating polyp-miss rates in recorded(colonoscopy)videos.

Methods A set of 95 anonymized white-light endoscopy-videos from a high-quality RCT were analyzed with CADe deep-learning(DL)system(GI Genius, Medtronic)that superimposes a green-square on colonoscopy images to direct the attention towards potential lesions. Although GI Geniusis engineered to work in real-time during live-endoscopy, a colonoscopy recorded without-AI can be fed to the device to annotate what would have been the highlighted areas.All the images containing at least one detection by GI Genius were recorded and reviewed by nine endoscopists,who classified each frame as a FP/TP. To investigate missed-lesions within the frames categorized as TP, subsequent frames classified as TP were then clustered into unique videoclips that were reviewed by three(expert)endoscopists, to exclude:(a)lesions excised later in the video;(b)suspect areas spotted by performing endoscopist, who decided to move over;(c)residuals from previous excision.

Results Overall(n=307)lesions were identified and resected in 95 videos(procedures). Revision of these videos by CADe-led to a set of 72 video clips of”candidate”missed lesions that were identified by at least 2/3 endoscopists, and a subset(n=28)of these candidates were considered as potential missed-lesions by all three expert endoscopists. The number of missed-lesions per patient measured by AI can thus be estimated between 0.29-0.76, while the corresponding miss rate(missed over missed plus resected lesions)is estimated between 8%-19%.

Conclusions In this study, we investigated for the first time the support of Artificial Intelligence(AI) in estimating polyp-miss rates in recorded colonoscopy videos. Measured miss rates are between 8%-19%. This represents a proxy for the contribution of failure in polyp recognition to such miss rate.