| Objective: In this study,LDCT was used as the gold standard and community screening chest DR film was used as the study object to investigate:(1)the diagnostic efficacy of AI for pulmonary nodules on chest DR film under different thresholds;(2)the diagnostic efficacy of AI on chest DR plain film compared with that of doctors in basic-level hospitals.Methods: Retrospective analysis of 199 patients who underwent chest DR from April2018 to July 2018 at the Community Hospital and underwent LDCT within 2 weeks at the Tianjin Medical University was performed.(1)Two associate chief doctors who have been engaged in chest imaging diagnosis labeled LDCT as the gold standard.(2)After the labeling was completed,the axial-labeled nodules were found on the coronal images by the Multi-Planner Reformation performed by three attending doctors who were engaged in chest imaging diagnosis.If the mark was in doubt,it should be determined through consultation.(3)The AI tag uses the Infer VISION’s Infer Read DR Chest Research R4.3,which used only chest DR frontal images,which were automatically fed into the system in batches,and the system automatically identified and marked lung nodules,the results of markup can be read directly through the visual interface.(4)The nodules on chest DR films in basic-level hospitals shall be marked by the above three attending doctors who are engaged in the diagnosis of chest DR films according to the registration form completed by the doctors in basic-level hospitals.Results:(1)LDCT marked 113 nodules above 6 mm.According to the different threshold values chosen,the nodules that the artificial intelligence recognized also had quite huge differences.Rate of missed diagnosis of AI increased with the increase of the threshold,but the false positive rate decreased with the increase of the threshold.The main factors affecting missed diagnosis rate and false positive rate were the overlap of nodules and bones,the location of nodules in Hilar region,the overlap of nodules and Heart Shadow,and other factors.Only the missed diagnosis caused by other factors(P=0.049/0.034<0.05)was statistically significant.It showed that the difference of missed diagnosis between different thresholds of the AI platform was mainly caused by other factors(mainly the limitation of the algorithm itself),and had no obvious correlation with the overlapping on chest DR film itself.The false positive caused by other factors(P=0.017/1.000)was statistically significant at low threshold,but not at high threshold.(2)Primary doctors diagnosed38 true nodules(sensitivity 33.6%).When the sensitivity of AI was 33.6%,the corresponding threshold was 0.585,and the false positive rate of AI was 9.When the false positive rate of AI was consistent with that of primary doctors,the corresponding threshold value is 0.525,at this time the artificial intelligence software diagnosis true knot is 44(the sensitivity is 38.9%).There was no significant difference between primary doctors and artificial intelligence software(the threshold value was0.585,0.525)in the influencing factors of missed diagnosis,the most important factor of both method was whether the nodules overlapped with the bone.The false positive caused by bone-related factors(P=0.010/0.008<0.05)was statistically significant,and the false positive caused by other factors(P=0.020/0.001<0.05)was also statistically significant.The results showed that when the nodules overlapped with the bones,the false positive nodules were more likely to be diagnosed in the overlapping regions by the doctors in basic-level hospitals.In contrast,artificial intelligence software was more likely to diagnose false positive nodules for other reasons.Conclusion:(1)Under different thresholds,the sensitivity and false positive rate of AI software are difficult to reach an ideal balance for nodules detected by chest DR plain film.No matter what the threshold value is,the skeletal related factors are all important factors that affect the rate of missed diagnosis and false positive rate,especially under the condition of low threshold value,at the same time,other factors in the high threshold condition become the most important factors affecting the missed diagnosis rate and false positive rate.(2)There is no significant difference in diagnostic efficiency between primary doctors and artificial intelligence software(threshold value was 0.585,0.525),and the factor that nodules overlapping with the bone could have a great impact on both of them. |
PDF Full Text Request