| ObjectiveBased on the three-step model of cervical cancer screening in China,this study focused on the feasibility of AI technology for three aspects of cervical cancer screening,triage and diagnosis.Initially,we explored the effectiveness of AI technology in cervical cytology screening,and considered whether it was feasible to improve cytology-based diagnostic screening performance and efficiency.Then,we assessed the clinical performance of the AI technology for triaging HPV-positive women,and attempted to identify the optimal triage strategy among the different algorithms included.Finally,we assessed the usefulness of AI technology in colposcopic diagnosis and biopsy,and we considered the implications for clinical practiceMaterials and Methods1.Feasibility evaluation of AI cytology in cervical cancer screening:This retrospective cross-sectional study involved 3,524 women from seven hospitals across China.We performed three tests including conventional cytology reading,AI reading,and AI-assisted reading,to assess the value of AI cytology in cervical cancer screening.Cytology expertise was taken as the gold standard,and the diagnostic performance of AI reading was assessed to investigate its generalization.AI analyzed each slide and generated risk scores which were used to determine the optimal configuration for algorithmically triaging true negative cases.We also evaluated whether cytologists’performance and efficiency could be improved using AI assistance.2.Evaluations of triage efficacy of AI cytology in HPV-positive population:Based on the cross-sectional study,489 HPV-positive women were included and triaged using AI cytology,liquid-based cytology and HPV16/18 test.Histologically confirmed cervical intraepithelial neoplasia grade 2/3 or higher(CIN2+/CIN3+)were considered disease thresholds for assessing clinical efficacy and exploring the optimal triage strategy.Additionally,pre-test and post-test probabilities(PPP)were estimated to reveal the utility.3.Application value of AI colposcopy in the detection of cervical lesions:Anonymized digital records from 19,435 patients in six hospitals across China.These records included colposcopic images,clinical information,and histological results(i.e.,the gold standard).Data were randomly assigned at a ratio of 7:1:2 to a training and a tuning set for developing AI colposcopy,or to a validation set for evaluating performance.Results1.Feasibility evaluation of AI cytology in cervical cancer screening:Standalone AI achieved 89.4%sensitivity and 66.4%specificity.The best sensitivity of AI cytology was 94.7%,and the best specificity was 79.5%in different hospitals across China.0.35 was established as the optimal AI configuration maximized triaging.1,319 slides were triaged without missing a single positive case.This also reduced cytologists’ workload by 39.6%(1,319/3,335).Reader analysis found that AI assistance had superior sensitivity and specificity compared to junior cytologists(81.6%vs 53.1%,78.9%vs 66.2%,respectively;all p<0.001).For senior cytologists,the specificity of AI assistance increased only marginally from 89.9%to 91.5%,(p=0.029).Sensitivity did not significantly increase from 84.4%to 87.2%,(p=0.45).2.Evaluation of triage efficacy of AI cytology in HPV-positive population:The sensitivity of AI cytology was comparable to cytologists(86.5%vs 83.8%,p=0.744),but substantially higher than HPV16/18 at detecting CIN2+(86.5%vs 54.1%,p=0.002).While the specificity of AI cytology was significantly lower than HPV16/18(51.3%vs 87.2%,p<0.001),significantly higher than cytologists at detecting CIN2+(51.3%vs 40.9%,p<0.001).The sensitivity of HPV16/18 with cytologists’ triage was 91.9%and therefore higher than cytologists alone.However,specificity was 36.7%which was lower than cytologists.HPV16/18 with AI cytology also achieved similar sensitivity with 91.9%,but showed a higher specificity with 46.0%.PPP plots suggested that if two-step triaging returned negative results,the risk of CIN3+was less than 0.55%,which falls into a "safer" green area.3.Application value of AI colposcopy in the detection of cervical lesions:The agreement between AI-graded colposcopic impressions and histological findings which was higher than colposcopies interpreted by colposcopists(82.2%versus 65.9%,kappa 0.750 versus 0.516,p<0.001).For detecting CIN2+,AI colposcopy showed higher sensitivity than colposcopists at either biopsy threshold(low-grade or worse:90.5%versus 83.5%;high grade or worse:71.9%versus 60.4%,all p<0.001).Specificities were similar(low-grade or worse:51.8%versus 52.0%;high grade or worse:93.9%versus 94.9%,all p>0.05.AI colposcopy also demonstrated a superior ability in predicting biopsy sites,with a median mean-Intersection-over-Union of 0.758.Conclusions1.Feasibility evaluation of AI cytology in cervical cancer screening:AI cytology had a high sensitivity and acceptable specificity,thus could be used for cervical cytology screening.This could reduce cytologists’ workload by more than one-third while improving diagnostic accuracy,especially when assisting less experienced cytologists.2.Evaluation of triage efficacy of AI cytology in HPV-positive population:AI cytology has equivalent sensitivity and higher specificity with more efficient colposcopy referrals for HPV-positive triaging,compared to cytologists.Combined triage algorithms for HPV16/18 with AI cytology could improve risk stratification of HPV-positive women,which is superior to HPV16/18 with cytologists.This could be particularly useful in settings where experienced cytologists are few in number.3.Application value of AI colposcopy in the detection of cervical lesions:Compared to colposcopists,AI colposcopy has higher agreement with histopathologic analysis,but higher sensitivity and equivolent specificity.AI colposcopy has potential in assist beginners and improve the diagnostic quality of colposcopy and biopsy in cervical precancer/cancer detection.Further research based on algorithm,intelligent grading,and human-computer interaction is necessary.External validation in real-world environments is also necessary to assess the long-term impact.To explore new AI models,suitable for cervical cancer screening,we must improve the efficiency of cytology screening and HPV-positive triaging,as well as colposcopy diagnosis and biopsy. |
PDF Full Text Request