Feasibility Evaluation Of Artificial Intelligence Electronic Colposcope Aided Diagnosis System

Objective:Cervical cancer is serious among female malignant tumors threating to women’s health.Early detection and treatmentcan control and reduce its occurrence effectively.The Chinese government attaches great importance to cervical cancer prevention and control and have been implemented across the whole nation.Colposcopic evaluation plays an important role in cervical cancer prevention,but it is greatly affected by doctors’expertise and clinical experience which lead to miss and over diagnosis in detection of cervical lesions.In light of this,we develop an artificial intelligence electronic colposcope assisted diagnosis system for assisting grassroots doctors to improve diagnostic level and efficiency.We also evaluated its performance in detecting cervical lesions,and discussed its potential application prospect.Methods:from November 2014 to December 2017,10000 patients who came to department of Shenzhen maternal and child health care hospital for cervical examination were selected to develop an artificial intelligence electronic colposcope assisted diagnosis system.The data were divided into learning sets(8000cases)and test sets(2000cases),according to the ratio of 80:20.Histopathological diagnosis was taken as the gold standard.The results of the system were compared with colposcopy doctors.Results:with histopathology as the gold standard,the accuracy rate of artificial intelligence colposcopy assisted diagnosis system for cervical lesions was 80.4%,and the accuracy rate of colposcopy doctors for cervical lesions was 71.2%.The difference was significant(x~2=46.141,P<0.05).The sensitivity,specificity,youden index and kappa value were 91.2%,85.8%,88.0%,0.75,respectively.The positive predictive value and negative predictive value of high grade squamous intraepithelial lesion(HSIL)were80.4%and 94.0%,respectively.The accuracy of colposcopists in diagnosing cervical lesions was 71.2%.The sensitivity of colposcopists was 75.9%,the specificity was85.0%,the youden index was 0.609,the coincidence rate was 81.5%,the kappa value was 0.61,the positive predictive value was 76.4%,and the negative predictive value was 84.6%.Further analysis of the accuracy of colposcopy diagnosis showed that the sensitivity of artificial intelligence electronic colposcopy assistant diagnosis system and colposcopy doctor was 91.2%and 75.9%,respectively,with significant difference(x~2=14.734,P<0.05);the specificity was 85.8%and 85.0%,with no significant difference(x~2=2.154,P>0.05),the positive predictive value of artificial intelligence electronic colposcope assisted diagnosis system was higher than that of colposcope doctors.Conclusion:the diagnostic accuracy,sensitivity and positive predictive value of the artificial intelligence electronic colposcope assisted diagnosis system for detecting high-grade lesions are high,thus it is expected to help clinicians to identify patients more accurately,improve the diagnostic ability of colposcopy,reduce the missed diagnosis caused by false negatives and the tissue damage caused by blind biopsy caused by false positives Compared with the diagnosis ability of colposcopy doctors,the artificial intelligence electronic colposcopy assisted diagnosis system has higher sensitivity and accuracy for detecting cervical lesions,which could improve the detection rate of CIN2+in cervical cancer screening and provide a solid guarantee for early screening and prevention of cancer at the grass-roots level.