Application Of Artificial Intelligence In Monitoring The Withdrawal Of Colonoscopy

Background and aims:Colorectal cancer(CRC)is the third most common cancer in the world.Colonoscopy can detect and remove precancerous lesions such as adenomas to reduce the incidence and mortality of CRC.Therefore,colonoscopy has been considered to be the most effective method for CRC screening.However,the quality of colonoscopy varies widely,and there is a certain rate of polyp missed rate.The withdrawal time is one of the important quality control indicators.Monitoring colonoscopy withdrawal time can improve adenoma detection rate and thus improve the examination quality,but withdrawal time cannot reflect the effective examination time of each segment of the colorectum,and cannot intervene in real-time because the documentation and feedback is relatively lagging behind.We developed a real-time colonoscopy withdrawal monitoring model based on artificial intelligence to achieve real-time monitoring of withdrawal stability and improve the efficiency of colonoscopy.This study was carried out in two parts:the development and verification of the colonoscope withdrawal stability real-time monitoring model and the exploration of appropriate withdrawal speed for polyp detection.Method:The first part was the construction and verification of the artificial intelligence-based real-time withdrawal stability monitoring model.The average hash algorithm and Hamming distance were used to calculate the similarity,and MobileNets was used to realize the recognition of blurred frames in the colonoscopy withdrawal video,then real-time colonoscopy withdrawal stability monitoring model was initially established.The complete colonoscopy videos from April 2019 to October 2020 was extracted from the database of the endoscopy center,Qilu Hospital of Shandong University.After segmentation and labeling,a total of 157 short videos were used to build the model:123 segments in the training set and 34 segments in the test set.And 3 complete colonoscopy videos were used for prospective verification.The second part was to explore suitable withdrawal speed for polyp detection.The instantaneous withdrawal speed before detecting the polyp was tested and compared with the average withdrawal speed.The relationship between the withdrawal speed and the characteristics of polyps was further analyzed.Results:1.Development and verification of the colonoscope withdrawal stability real-time monitoring model:The performance of the model was as follows:the sensitivity of stable withdrawal was 77%,the specificity was 73%,the accuracy was 74%,the positive predictive value was 62%,and the negative predictive value was 84%.The sensitivity of unstable withdrawal was 73%,the specificity was 77%,and the accuracy was 74%,the positive predictive value was 84%,and the negative predictive value was 62%.The area under the curve of stable withdrawal was 0.82,and the area under the curve of unstable withdrawal was 0.82.The recall rate of stable withdrawal was 0.77,the precision rate is 0.84,and the F1 score is 0.78.The recall rate of unstable withdrawal was 0.73,the precision rate is 0.62,and the F1 score is 0.69.verification results:AI detected a total of 37 unstable withdrawal,endoscopist marked a total of 30 unstable withdrawal,there is a moderate degree of consistency between AI and endoscopist,Kappa value of 0.718,p<0.001.2.The exploration of appropriate withdrawal speed for polyp detection,analysis of withdrawal speed and polyp characteristics:A total of 246 short polyp positive colonoscopy videos of 104 patients were included in the analysis to test the instantaneous withdrawal speed,and the average and instantaneous endoscopic withdrawal speed were compared.The results showed that the instantaneous withdrawal speed was slower than the average withdrawal speed,6(3.5,12)VS 17(14,20),P<0.001.Further analysis of the instantaneous withdrawal speed of polyps showed that the withdrawal speed of proximal colonic polyps was slower than that of distal colorectal polyps 5(3,8.5)VS 6(4,14),P=0.039;The detection speed of flat polyps was slower than that of.sessile colorectal polyps,4(3,9)VS7(4,13),P=0.013.Conclusions:1.The system has high performance in real-time monitoring the stability of the colonoscopy withdrawal,and has a moderate agreement with the endoscopist.2.The instantaneous withdrawal velocity when polyps were detected was often lower than the slow velocity observed subjectively and average withdrawal velocity.3.The detection of flat polyps and polyps in the proximal colon requires a slower speed of withdrawal.