| There are estimated 1.88 million new colorectal cancer(CRC)cases diagnosed,and 0.92 million CRC caused deaths in 2020 worldwide.CRC has seriously threatened human life and health.Colonoscopy has attained a prominent place in preventing CRC.Endoscopists can detect and remove polyps that are pathologically confirmed as adenomas during colonoscopy,consequently diminishing the risk of CRC.However,on the one hand,there are substantial differences in the quality of colonoscopy,and a considerable number of adenomas can be missed.On the other hand,the pathological analyses for diagnosing the adenoma will increase medical costs.The time taken for pathology analysis prevents the endoscopist from deciding which treatments are available to patients in real time.As a result,the patient needs to be re-examined.Deep learning has achieved excellent results on multiple complex image classification tasks,matching or even surpassing human performance.In this study,deep learning was used to construct a quality control and polyp classification system from three aspects: realizing a complete colonoscopy,prolonging the withdrawal time and improving accuracy in optical diagnosis of colorectal polyps.Section one: Development and validation of a real-time quality control and polyp classification system based on deep learningObjective: To development a real-time colonoscopy quality control and polyp classification system based on deep learning,which is used to record the insertion and withdrawal time,monitor the speed of withdrawal in real time,and determine whether colorectal polyps are adenomatous or non-adenomatous during colonoscopy.Methods: 41,982 in vivo images and 12,220 in vitro images were collected to train and test model A.5,223 clear cecum images and 5,685 clear non-cecum images were collected to train and test model B.5,148 images of narrow-band imaging(NBI)and 5,016 images of white light were collected to train and test model C.2,987 adenomatous polyp images and 2,056 non-adenomatous polyp images were collected,and model D was trained and tested.Based on Resnet50 algorithm,model A-D were constructed.Model A and model B were used for recording insertion and withdrawal time.Model C and Model D were used for polyp classification under NBI.Perceptual hashing algorithm and Hamming distance were used to calculate the similarity of adjacent pictures to reflect the withdrawal speed.After model construction,videos were prospectively collected to validate the performance of the system.Results: Model A,model B,model C,and model D were tested using images.The accuracy of model A-D were 99.87%(95% CI,99.73%-99.94%),96.70%(95%CI,95.47%-97.61%),99.71%(95% CI,99.15%-99.90%)and 90.16%(95% CI,87.23%-92.48%).In the videos,the accuracy of the system in recording insertion and withdrawal time was 95.00%(19/20).Among the 12 colonoscopy videos labeled by endoscopists,the average withdrawal speed of the faster withdrawal clips was significantly higher than that of the slower withdrawal clips(22.62 vs 13.25,P <0.001).Conclusions: The colonoscopy quality control and polyp classification system based on deep learning can record the insertion time and the withdrawal time,monitor the withdrawal speed in real time,and determine whether colorectal polyps are adenomatous or non-adenomatous under non-magnifying NBI.The system has good performance in images and videos and is expected to improve the quality and cost-effectiveness of colonoscopy in future clinical work.Section two: Study on colonoscopy withdrawal speed monitoring systemObjective: Prolonging the withdrawal time can improve the adenoma detection rate(ADR).However,the feedback of the withdrawal time is relatively lagging.Therefore,we constructed a withdrawal speed monitoring system which can enable endoscopists to withdraw the scope at a steady and slow speed,so that each colon segment can be fully observed.The purpose of this study was to further explore the correlation between withdrawal speed and ADR,and to find an appropriate withdrawal speed during colonoscopy.Methods: A retrospective observational study was conducted at the Digestive Endoscopy Center,Renmin Hospital of Wuhan University.Patients who underwent colonoscopy from May 2020 to September 2020 were included.The videos recording of their colonoscopy procedure were collected,and the perceptual hash algorithm and Hamming distance were used to calculate the withdrawal speed.The relationship between the withdrawal speed and the ADR,polyp detection rate(PDR)and withdrawal time were analyzed and evaluated using spearman correlation.Then,multivariate logistic regression analysis was performed to explore the risk factors affecting the ADR.Results: 493 patients were enrolled,98 of whom had adenoma.The ADR was19.88%.There was a significant negative correlation between the withdrawal speed and the ADR(r=-0.905,P<0.01).There was a significant negative correlation between the withdrawal speed and the PDR(r=-0.738,P=0.037).There was a significant negative correlation between withdrawal time and withdrawal speed(r=-0.476,P<0.001).Among 389 patients undergoing screening colonoscopy,when the withdrawal speed was 10-17,the corresponding ADR in each group was >20.00%.According to the index of ADR>20% proposed in the guidelines,we defined the withdrawal speed of <18 as the safe withdrawal speed.The results of multivariate logistic regression analysis showed that adenomas were more likely to be detected when the withdrawal speed was less than 18 than when the withdrawal speed was greater than or equal to 18(OR: 0.526;95% CI,0.310-0.893).Conclusion: The withdrawal speed monitoring system can quantitatively analyze the withdrawal speed of colonoscopy,and there is a significant negative correlation between withdrawal speed and ADR、PDR and withdrawal time.Withdrawal speed is an independent factor affecting the ADR.Performing colonoscopy at the safe withdrawal speed defined in this study appears to promote the quality of colonoscopy.Section three: Application of deep learning in improving the accuracy of colorectal polyp classificationObjective: Narrow-band imaging(NBI)can assist endoscopists in observing the microstructure of the intestinal mucosal epithelium more clearly,and predict whether the polyp is adenomatous or non-adenomatous.Deep learning was used to construct a system for colorectal polyp classification in real time,aiming to assist endoscopists in diagnosis and improve their performance of polyp classification.Methods: Non-magnifying NBI polyp images from endoscopy center of Renmin Hospital of Wuhan University were collected(2,699 adenomatous and 1,846non-adenomatous polyp images from January 2018 to October 2020)for model training and validation.288 adenomatous and 210 non-adenomatous polyp images from January 2018 to October 2020 was used to compare the accuracy between the system and endoscopists.208 adenomatous and 141 non-adenomatous polyp images were prospectively collected to test the model.Then,the accuracy of 8 trainees in polyp classification with and without the assistance of this system was compared.Results: The accuracy of the system in polyp classification was 90.16%(95% CI,87.23%-92.48%)which was better than that of all endoscopists.In the prospective test,the accuracy of the system was 89.53%(95% CI,85.85%-92.34%).With the assistance of the system,the novice endoscopists’ accuracy in polyp classification was significantly improved(78.22% vs 86.71%,P <0.001)and the diagnosis time was shortened from 5.60 to 4.30s(P =0.012).Conclusion: The colorectal polyp classification system based on deep learning developed in this study can significantly improve the ability of trainees in polyp classification and shorten the diagnosis time. |
PDF Full Text Request