| Background and aims:Colorectal cancer(CRC)is a common malignant disease worldwide.In China,there are approximately 370,000 new cases of CRC and 190,000 related deaths each year,making it the fifth leading cause of cancer deaths.Most CRCs arise from adenomatous polyps,which take more than 10 years on average to progress to cancer.For this reason,effective screening can prevent the development of CRC.Colonoscopy as the gold standard for CRC screening,can detect and remove polyps early,thus significantly reducing CRC morbidity and mortality.However,due to many influential factors,the overall miss rate of colorectal polyps in conventional colonoscopy is as high as 22%.Therefore,the development of a new technology,artificial intelligence(AI)-assisted diagnosis,to improve polyp detection in colonoscopy,has become a new research hotspot in the field of endoscopy.Based on these,our study of AI-assisted polyp detection for colonoscopy was carried out in three steps:methodological exploration,new system development and preliminary validation,and a real-time application clinical trial This study aimed to develop a new AI-assisted polyp detection system and confirm it effectiveness in real-time application,providing an effective strategy for polyp detection in colonoscopy.Methods:The first part of our study was the methodological exploration of AI-assisted polyp detection for colonoscopy.Using a deep learning method,we initially constructed an AI-assisted polyp detection system for colonoscopy based on the Faster R-CNN algorithm.Five test groups were set according to the size and difficulty of the training set:test group 1,2,3,and 4 contained 1,000,2,000,4,000,and 6,000 training samples respectively,and test group 5 contained 6,000 training samples with increased probability of selecting difficult samples.The testing sets for all test groups were the same.For each test group,the image classification indicators(such as sensitivity,specificity,etc.)and target detection indicators(such as recall,precision,etc.)of this system were evaluated.The second part of the study was the preliminary validation of the AI-assisted polyp detection system for colonoscopy.Based on the methodological exploration in the first part,the polyp detection system used in this part is a new Al-assisted real-time polyp detection system developed on the RestinaNet network model,using a deep learning method.The self-controlled study was conducted to confirm the effectiveness of this system by comparing the polyp detection results of colonoscopy and AI system in the same population.The primary outcome in this part of study is polyp detection rate(PDR).Secondary outcomes include polyps per colonoscopy(PPC)and polyps per colonoscopy-plus(PPC-Plus).The third part of the study was a real-time application clinical trial of AI-assisted polyp detection for colonoscopy.The system used here is the newly developed AI-assisted real-time polyp detection system in the second part.To evaluate the safety and efficacy of this system in real-time colonoscopy,we conducted a prospective,multicenter,randomized controlled trial to compare the polyp detection results of the conventional colonoscopy group and the AI-assisted colonoscopy group.The primary outcome in this part of study is PDR.Secondary outcomes include PPC and PPC-Plus,as well as polyps per positive patients(PPP).Results:In the methodological exploration study,an AI-assisted colonoscopy polyp detection system based on the Faster R-CNN algorithm was initially constructed using real colonoscopy images from six endoscopy centers as the study materials(a total of 10,061 colonoscopy images of 5,844 patients with colorectal polyps,of which 6,000 were used as training set and 4,061 were used as testing set).Comparing the system testing results under different training sets in the deep learning process,the image classification results showed that the sensitivities of test group 1-5 were 90.1%,93.3%,93.3%,93.3%and 93.5%,respectively,and the difference between them was significant(P<0.001),as the sensitivities of test group 2-5 were significantly higher than that of test group 1(Ps<0.00625).There were no significant differences in specificity and positive predictive value between the test groups;while the negative predictive values were significantly different(P<0.001),as the negative predictive values of test group 2-5 were significantly higher than that of test group 1(Ps<0.00625).According to the receiver operating characteristic curve,when the training sample size is 1,000,the AUC value is 0.941.When the sample size is increased to 2,000,the AUC is increased by 0.02,but when further increasing the sample size to 6,000,the AUC is increased by no more than 0.01.Then,increasing the training difficulty without changing the sample size,the AUC increased by 0.004,reaching 0.973.In addition,the target detection results showed that the recall rates of test group 1?5 were 73.6%,79.8%,79.5%,79.8%and 83.3%,respectively,and the difference between them was significant(P<0.001),as the recall rates of test group 2-4 were significantly higher than that of test group 1(Ps<0.00625),and the recall rate of test group 5 was significant higher than those of test group 1-4(Ps<0.00625).The precision rates between the test groups were significant different(P<0.001),as the precision rates of test group 3 and 5 were significantly higher than that of test group 2(Ps<0.00625),and the precision rate of test group 4 is significantly higher than test group 1 and 2(Ps<0.00625).As the training sample size and difficulty increased,the F1 score and mean average precision(mAP)gradually increased.Based on the experience in methodology exploration,we newly developed an AI-assisted real-time polyp detection system based on the RestinaNet network model.It differs from the previous system in that the new system enables real-time endoscopic image analysis and improves the detection of diminutive polyps and even incomplete imaging polyps.Using a large number of real colonoscopy videos(total 117,048 frames)for system training and testing,the system recall rate and precision rate reached 81.9%and 93.4%,respectively.In the self-controlled study,a total of 764 subjects were included for polyp detection.The PDR of the AI system was significantly higher than that of conventional colonoscopy(45.5%vs 35.5%,P<0.001),and its PPC(1.1 vs 0.7,P<0.001)and PPC-Plus(0.6 vs 0.4,P<0.001)were also significantly higher.In terms of polyp characteristics,the location distribution of the polyps detected by the AI system was similar to that of conventional colonoscopy,but more diminutive polyps and flat polyps were detected by AI.Further subanalysis of PDR for different levels of colonoscopists showed that the PDR of the AI system was significantly higher than that of conventional colonoscopy for both junior operators(43.9%vs 34.2%,P<0.001)and senior operators(47.6%vs 37.0%,P<0.001).A total of 2,352 subjects were included in the randomized controlled trial of real-time application of the AI-assisted polyp detection system,including 1,175 in the conventional colonoscopy group and 1,177 in the AI-assisted colonoscopy group.No adverse events associated with the use of the system occurred during the study.There were no significant differences between the two groups in terms of general clinical characteristics(gender,age,BMI and waist circumference),chief complaints,and colonoscopy related control indicators(anesthesia,BBPS score,cecal intubation time and withdrawal time).The PDR of the AI-assisted group was 2.6%higher than that of the conventional group(38.8%vs 36.2%,P=0.183),and the PPC-Plus was significantly higher than that of conventional group(0.5 vs 0.4,P<0.05).PPP and PPC also showed an increasing trend.In terms of polyp characteristics,the AI-assisted colonoscopy detected more type IIa polyps than conventional colonoscopy.Further analysis of the polyp biopsy results revealed that the pathological type composition of polyps detected by AI-assisted colonoscopy was similar to conventional colonoscopy,both detecting more adenomatous polyps.In addition,a subanalysis by endoscopic center showed that the PDR of the AI-assisted group,as well as PPC and PPC-Plus,were significantly higher than those of conventional group at Zhejiang University Ningbo Hospital.In Yuyao People's Hospital of Zhejiang Province and Sanmen People's Hospital,the PPP and PPC-Plus of the AI-assisted group were significantly higher than conventional group;there were no significant effects in other endoscopic centers.Conclusions:The methodological exploration study helped initially construct an AI-assisted polyp detection system for colonoscopy.As the training sample size and difficulty increased,the system performance could be improved.Finally,the sensitivity and recall rate reached 93.5%and 83.3%,respectively.In the self-controlled study,our newly developed AI-assisted real-time polyp detection system can significantly improve polyp detection in colonoscopy,finding more positive patients and more colorectal polyps.It is more effective in detecting diminutive polyps and flat polyps.In addition,the improvement in polyp detection with the AI system is significant for any level of colonoscopists,thus reducing the operator dependence.The real-time application of the AI-assisted polyp detection system can improve colorectal polyp detection to some extent through increasing the polyp detection rate by 2.6%and significantly increasing the detection of non-first polyps and slightly elevated polyps.No adverse events associated with the use of the system occurred during the study.These prove that the real-time application of the AI-assisted detection system for polyp detection is safe and effective.In summary,based on the methodological exploration,new system development and preliminary validation,and a real-time application clinical trial,this study confirmed the safety and effectiveness of our newly developed AI-assisted polyp detection system for colonoscopy.It provides evidence to support further clinical application of this system. |
PDF Full Text Request