The Construction Of Electronic Colposcopic Auxiliary Diagnostic System Based On Artificial Intelligence And Multicenter Application Evaluation

Objectives1.To build Colposcope Artificial Intelligence Auxiliary Diagnosis System(CAIADS),and evaluate diagnostic ability for detecting cervical lesions and guiding cervical biopsy sites in a hospital-based multicenter retrospective study using pathological diagnosis as the gold standard.2.The performance of the CAIADS in guiding cervical biopsies are compared with the colposcopists of three expertise levels(expert,competent and junior).Materials and MethodsThis retrospective,multicentric,diagnostic study was done in six hospitals of different tiers(municipal and provincial)in China.In December 2018,anonymized digital records of patients,including colposcopic impressions,images,cytology,human papillomavirus(HPV)testing,and pathological results as the gold standard were retrospectively collected from Shenzhen Maternity and Child Healthcare Hospital(SZMCHH)for training and internal validation of the CAIADS.To further ascertain diagnostic performance of the CAIADS for detecting cervical lesions,external validations were conducted in five municipal or provincial hospitals across China:Chongqing University Cancer Hospital(CQUCH),Chengdu Women’s and Children’s Central Hospital(CDWCCH),Jiangxi Maternity and Child Health Hospital(JXMCHH),Liaoning Cancer Hospital(LNCH),and Henan Cancer Hospital(HNCH).Moreover,the performance of the CAIADS was compared with the colposcopists with three expertise levels(expert,competent and junior).In statistic analysis,the Chi-square test was used to compare the inter-group rates,and a was set to 0.05.Results1.The performance of the CAIADS for detecting cervical lesions:In the internal validation set,diagnostic accuracy,sensitivity and specificity of the CAIADS for detecting cervical lesions were 0.954(95%CI:0.941-0.964),0.957(95%CI:0.939-0.972),and 0.950(95%CI:0.932-0.965),respectively.In the five external validation sets,the accuracy,sensitivity and specificity all exceeded 0.850.2.The performance of the CAIADS in guiding cervical biopsy sites:by comparison with the location of biopsy sites indicated by pathological results,we found that a median Intersection-Over-Union(IOU)of 0·705 was achieved by the CAIADS.The higher IOU value represented the more accurate performance for cervical biopsy site prediction.3.Comparison of diagnostic ability of the CAIADS with the colposcopists of three expertise levels for detecting cervical lesions:Based on all the participants of validation sets from our study,the CAIADS showed lower diagnostic accuracy,sensitivity,and specificity compared with expert colposcopists with over 10 years of clinical experience,(0.886[0.877-0.894]vs.0.947[0.935-0.957],p<0.001;0.902[0.890-0·914]vs.0.981[0.969-0.989],p<0.001;0.873[0.861-0.885]vs.0.915[0.896-0·932],p<0.001);but similar accuracy,sensitivity,and specificity compared with competent colposcopists with 5-10 years of clinical experience(0.877[0.860-0.892],p>0.05;0.882[0.858-0.903],p>0.05;0.872[0.848-0.894],p>0.05)and significantly higher accuracy,sensitivity,and specificity compared with junior colposcopists with 3-5 yeas of clinical experience(0.546[0.511-0.580],p<0.001;0.774[0.725-0.818],p<0·001;0.392[0.348-0.436],p<0.001).Conclusions1.We built a Colposcope Artificial Intelligence Auxiliary Diagnosis System(CAIADS)using previous datasets from cervical outpatient clinics.The CAIADS achieved an excellent performance for detecting cervical lesions in the internal validation set and additional five external validation sets,which could be used for guiding cervical biopsies.2.Diagnostic accuracy,sensitivity and specificity of the CAIADS were observed inferior to expert colposcopists,but similar to competent colposcopists and significantly superior to junior colposcopists.It could assist the colposcopists to improve biopsy quality.Based on the CAIADS,we could create a new cervical cancer screening model to improve clinical colposcopic diagnosis process,and build a cloud service platform without regional restrictions,even in remote areas which can upload colposcopic images to the network for real-time,intelligent and automatic diagnosis.The development of CAIADS is expected to solve current bottlenecks of colposcopy diagnosis,thereby improving cervical cancer screening performance of China.