Objective To set up a computer-assisted polyp detection system under colonoscopy,and to preliminarily verify its effectiveness.Methods Based on Faster R-CNN algorithm and the open source implementation of the open source framework tensorflow and Faster R-CNN,a computer-assisted polyp detection system under colonoscopy was constructed.According to the size and difficulty of the training set,five test groups were set up:test group one,two,three and four contained 1 000,2 000,4 000 and 6 000 training samples,respectively.Test group five increased the probability of selecting the difficult samples based on 6 000 training samples.In different training sets,the sensitivity,specificity,other classification evaluation parameters,and the evaluation parameters of target detection such as recall and precision of this polyps detection system were calculated.Results Classification evaluation parameters showed that the sensitivities of test group one,two,three,four and five were 90.1％,93.3％,93.3％,93.3 ％ and 93.5 ％,respectively,and the difference was statistically significant (x2 =25.324,P＜0.01).The sensitivities of test group two,three,four and five were all higher than that of test group one,and the differences were statistically significant (x2 =13.964,13.508,13.508 and 13.386,all P＜ 0.006 25).There were no significant differences in specificity and positive predictive value among test groups (both P＞0.05).The negative predictive values of test group one,two,three,four and five were 90.4％,93.3％,93.3％,93.3％ and 93.5％,respectively,and the differences were statistically significant (x2 =21.862,P＜0.01).The negative predictive values of test group two,three,four and five were higher than that of test group one,and the differences were statistically significant (x2=11.447,11.564,11.755,13.760;all P＜0.006 25).As the training sample size increased from 1 000 to 2 000,the area under curve (AUC) increased by 2％,and further increased the sample size to 6 000,AUC increased by less than 1 ％.At this point maintaining the same sample size while increasing the proportion of difficult samples,AUC increased by 0.4％.The results of evaluation parameters of target detection showed that the recall rate of each test group was 73.6％,79.8％,79.5％,79.8％ and 83.3％,respectively,and the differences were statistically significant (x2 =71.936,P＜0.01).Among them,the recall rates of test group two,three and four were higher than that of test group one,and the differences were statistically significant (x2 =25.960,23.492 and 25.960,all P＜0.006 25),and the recall rate of test group five was higher than those of test group one,two,three and four,and the differences were statistically significant (x2=67.361,9.899,11.527 and 9.899;all P＜0.006 25).In addition,the precision rates of test group one,two,three,four and five were 87.9％,85.3％,90.2％,91.4％ and 89.2％,respectively,and the difference was statistically significant (x2=48.194,P＜0.01).The precision rates of test group three and five were higher than that of test group two,and the differences were statistically significant (x2 =24.508 and 15.223,both P＜0.006 25),and the precision rate of test group four was higher than those of test group one and two,and the differences were statistically significant (x2=13.524 and 39.120,both P＜0.006 25).As samples size and training difficulty increased,the values of F1-score and mean average precision increased steadily.Conclusions This study initially constructed a computer-assisted polyp detection system under colonoscopy.Currently the maximum sensitivity reached 93.5％,and the maximum recall rate reached 83.3％.Increasing the training set size may improve the polyp detection result to a certain degree,however it will reach a bottleneck.At this time,increasing the training difficulty can further improve the detection scores,especially the recall rate.