Predictive Model For Endoscopic And Pathological Classification Of Pre-Cancerous Colon Polyps Based On Deep Convolutional Neural Network

Objective: CRC is one of the most common malignancies worldwide.Most CRC development follows an adenoma-carcinoma evolution,but the diagnosis of colon polyps requires extensive experience of endoscopists.In order to reduce the risk of CRC,an intelligent diagnostic model is needed to assist diagnosis of colorectal polyps,whose performance is close to experienced endoscopists.The study aims to combine CNN with colonoscopy to establish an intelligent diagnostic model that can both accurately identify colon polyps and perform endoscopic and pathological classification prediction for polyps.This model provides recommendations for endoscopists in the diagnosis of colonic polyps.Methods: This study included 2334 pictures of patients who underwent colonoscopy at the Surgical Endoscopy Center of the First Hospital of Lanzhou University from July 2021 to January 2023.These included images of white light,NBI,magnified endoscopic colon polyps,and images of normal intestinal cavity.These images were randomly divided into training and test set.Based on the CNN algorithm,the training set was used to construct an endoscopic and pathological classification prediction model for colon pre-cancerous polyps,and the test set were used for validation.The model was compared with the classification results of experienced endoscopists to evaluate the diagnostic effectiveness of the predictive model.Results: 1.In this study,about 2104 pictures of colon polyps under white light,NBI and magnifying colonoscopy and 230 pictures of normal intestinal cavity were included,including 1900 pictures of colon polyps in training set and 204 pictures of colon polyps in test set.In the test set,Pit Pattern subtypes IIIL and IV had an accuracy of 91.77%and 98.26% in this prediction model,with a specificity of 92.70% and 99.10%.The accuracy and specificity of type 2 in NICE were 90.04% and 84.88%.The accuracy and specificity of type 2A and type 2B in JNET classification were 87.87%,96.10%,79.26%and 99.54%.The accuracy and specificity of the pathological classification of adenomatous polyps in the model was 91.77% and 95.10%.The areas under ROC curve of Pit Pattern classification,NICE classification,JNET classification and pathological classification were 0.87,0.96,0.96 and 0.83,respectively.2.There was good concordance between the AI group and the endoscopist group in the assessment of the classification of colon polyps.The Kappa value of Pit Pattern was0.77(95% CI,0.64-0.90).Kappa value of NICE was 0.56(95% CI,0.38-0.74).JNET Kappa value was 0.61(95% CI,0.45-0.76).The Kappa value of pathological classification was 0.57(95% CI,0.36-0.78).Conclusion: In this study,based on the CNN model,the endoscopic and pathological classification prediction model for colon pre-cancerous polyps was constructed.The model can not only identify colon pre-cancerous polyps,but also it has good accuracy in endoscopic and pathological classification.The model is comparable to the diagnostic ability of senior endoscopists,but its stability and applicability need to be further validated.