Big Data Investigation Of Colorectal Polyps And Intelligent Detection Of Colorectal Polyps Guided By Geometry

Backgound Colorectal Cancer(CRC)is one of the most common malignant tumors in the digestive system.Colorectal polyps are the most important precancerous lesions of colorectal cancer.Colonoscopy is the gold standard for the diagnosis of colorectal cancer and polyps.The removal of polyps during colonoscopy can reduce the total mortality rate of colorectal cancer.However,professional doctors also have a certain degree of omission or misdetection in the judgment of images.In order to reduce the missed detection rate of polyps,Computer-Aided Diagnosis(CAD)attracts more and more attention from endoscopists.At present,CAD has made some progress in colorectal polyp detection technology,but the existing methods do not make full use of the natural timing and spatial information in colonoscopy video.Real-time and accurate polyp detection is performed,and the data set is small.Therefore,this study will combine the time and space related information of colonoscopy video to develop a real-time and effective computer-aided detection system for colorectal polyps with a high recall rate.Objective(1)To investigate the clinical characteristics and risk factors of adenomatous polyps and high-risk adenomas,as well as the clinical characteristics and risk factors of recurrent colorectal polyps after resection;(2)Establish a sound database of endoscopic images and videos in line with industry standards,develop a relatively complete computer-aided colorectal polyp detection system based on endoscopic videos,and conduct clinical tests through colonoscopy videos to verify its accuracy and reliability.Methods(1)The data of 5732 polyps in 4055 patients with colorectal polyps were retrospectively analyzed.According to the pathological results of polyps,they were divided into adenomatous polyps group(3547 cases)and non-adenomatous polyps group(2185 cases),low-risk adenoma group(3265 cases)and high-risk adenoma group(282 cases)to analyze the characteristics and risk factors of aden omatous polyps and high-risk adenoma.The clinical data of 370 patients who un derwent two colorectal polypectomy were retrospectively analyzed.According to the number of polyps in the second colonoscopy,the patients were divided into multiple recurrence group(≥3)and non-multiple recurrence group(< 3),and the characteristics and risk factors of polyp recurrence were analyzed.(2)70 colorectal endoscopic videos were clipped and segmented,and polyp image labeling was performed to construct a large data set with rich forms;The deep learning model of colorectal polyp intelligent detection was constructed by using the time sequence information contained in colonoscopy videos,and the model was tested by 93 colonoscopy videos in the video library.Results(1)Age(AOR=1.021,P < 0.001,),polyp size(AOR=2.380,P < 0.001),B BPS(AOR=1.083,P=0.002),polyp morphology(AOR=2.640,P < 0.001),polyp col or(AOR=0.162,P < 0.001),right colon(rectum as reference,AOR=1.354,P = 0.001),left colon(rectum as reference,AOR=1.318,P= 0.001)were independent risk factors for finding adenomatous polyps;Age(AOR=1.022,P < 0.001),polyp size(AOR=2.672,P < 0.001),resection method(AOR=4.121,P < 0.001),rectum(left colon as reference,rectal AOR=1.461,P =0.037),non-colonic pocket(non-colonic pocket as reference,Anal AOR=0.59,P=0.042,colon pocket AOR=0.627,P=0.002)were independent risk factors for finding high risk adenomas.In univariate stati stical analysis,polyp size(P=0.033),polyp site(P= 0.002)and polyp color(P=0.032)were the risk factors for multiple polyp recurrence.In multivariate regress ion analysis,the number of polyps first detected(P=0.001)and the location of polyps(P=0.001)were independent risk factors for multiple recurrence.(2)A large-scale endoscopic video database LDPolyp Video with 160 segments containing 40266 frames of images and 200 polyps was constructed.(3)The diagnostic accuracy of the model was 97.85%(95%CI 92.49-99.41),the specificity was 100%(95%CI 88.3-100),and Positive Predictive Value was 100%(95%CI 94.17-100),Negative Predictive Value 93.55%(95%CI 70.28-98.21),which was higher than that of the junior doctors and more consistent with the expert diagnosis results(Kappa=0.951,P<0.001).The detection rate of polyp was 91.608%,which was higher than that of the junior doctors group(P < 0.001).Polyp size(AOR=12.584,P <0.001)and polyp shape(AOR=22.701,P=0.03)were independent risk factors for inconsistencies between the junior physicians and the detection model.Conclusion(1)There are differences between adenomatous polyps and non-adenomatous polyps in patient age,polyp size,location,shape,color and intestinal cleanliness.There are differences between low-risk and high-risk adenomas in age,size,resection method,polyp site and colon pocket distribution.High-quality intestinal preparation is beneficial to detect adenomatous polyps.The risk of adenomatous polyps and high-risk adenomatous polyps with villous tissue increased with age and polyp size.The red and raised polyps were more likely to be adenomatous polyps.The right and left colon had a higher risk of adenomatous polyps compared with the rectum,and the rectum had a higher risk of high-risk adenoma than the left colon.Non-supramental(transverse groove)adenomas are at higher risk for adenomas than proximal anal and supramental pockets.(2)The number and location of polyps in the first colonoscopy are independent risk factors for multiple recurrence after polypectomy.The more polyps found and removed under colonoscopy,the more likely multiple recurrence of polyps will occur,and the more likely multiple recurrence of polyps will occur in the left colon.(3)A large-scale and complete nodal intestinal endoscopic image and video database conforming to industry standards,LDPoly Video,has been established and opened to the whole research community.(4)The computer-aided colorectal polyp detection model established in this study has high polyp detection efficiency,which can assist junior physicians to perform polyp detection,improve the polyp detection rate and reduce the rate of missed diagnosis.