Artificial Intelligence-assisted Interpretation Of Ki-67 Expression And Repeatability In Breast Cancer

Objective: Compared the repeatability of the visual evaluation under microscope(Visual Assessment,VA),reference breast cancer Ki-67 standard reference card(Ki-67 Standard Reference Card of Breast Cancer,SRC),manual counting under the microscope(Manual Counting,MC),and artificial intelligence automatic counting(Artificial Intelligence Counts,AI).Explore the clinical applicability and accuracy of artificial intelligence assisted interpretation of breast cancer Ki-67.Methods: Collecting 300 cases of modified radical mastectomy of breast cancer in the Fourth Hospital of Hebei Medical University from October 2017 to October 2019.150 cases were used for training and an other 150 cases were used for verification.Ki-67 immunohistochemical s taining sections were scanned and converted into image data using the PRECICE 600 series automatic digital section scanning system.Using T encent artificial intelligence software to make Ki-67 standard reference c ards ranging from 5% to 90%.In the training group,nine pathologists w ith different years of age independently performed the interpretation of Ki-67 immunohistochemistry through VA,SRC,MC,and AI under the s ame scene at two-week intervals.The results were analyzed by SPSS 23.0 statistical software.Kolmogorov-Smirnov(K-S)is used for normality test,Friedman M and Bonferroni correction are used for difference ana lysis.For consistency,Intraclass correlation coefficient(ICC)and Bland Altman scatter plot were used.Results:1 The training group1)The Ki-67 LI interpreted by the four methods did not obey the normal distribution(P<0.05).The difference between observers was statistically significant(P<0.05).After correction,the Ki-67 LI interpreted by SRC and AI only showed a statistically significant difference between the pathologist and the gold standard(P<0.05).There was no significant difference between pathologists(P>0.05).2)AI-assisted interpretation of Ki-67 LI was significantly better than that of human interpretation.The ICC of Ki-67 LI read by the nine pathologists through SRC were 0.918(0.899,0.936),and the ICC of AI were 0.972(0.964,0.978).The ICC of VA was 0.757(0.711,0.802),and the ICC of MC was 0.803(0.763,0.841).Through SRC and AI,the consistency between nine pathologists and the gold standard was very good,and ICCs were all above 0.80.3)Nine pathologists were divided into three groups according to their working years,three were senior,three were middle and three were primary.Comparing ICC before and after the SRC,the ICC of all three primary pathologists increased to above 0.80 after passing the SRC.4)In the homogeneous group,the consistency among observers of VA,MC,SRC and AI was very good,and the ICCs were all above 0.80.In the heterogeneous group,SRC and AI count Ki-67 LI showed good agreement among observers(ICC was 0.877(0.844,0.907)and 0.959(0.947,0.970)).And poor consistency among pathologists with VA and MC(ICC was 0.698(0.634,0.760)and 0.672(0.605,0.738))2 The verification groupBy SRC,the three pathologists have an ICC of 0.988(0.985,0.911).By AI,the ICC is 0.990(0.987,0.992).All three pathologists were in good agreement with the gold standard,with ICC values above 0.95.In the homogeneous group and the heterogeneous group,the consistency among SRC and AI pathologists was very good.ICC values were 0.985(0.978,0.990),0.986(0.979,0.990),0.990(0.986,0.993),and 0.991(0.986,0.994).Conclusions:1.AI significantly improves the repeatability between pathologists,and the true value is closer to the gold standard.It can be seen that AI interpretation is a good tool to improve the reproducibility of breast cancer Ki-67 LI.2.The Ki-67 standard comparison card increases the consistency of pathologists in interpreting breast cancer Ki-67 LI,especially for primary and middle pathologists.In the case that AI software has not yet become popular,the Ki-67 standard comparison card A simple tool to assist in the interpretation of breast cancer Ki-67 LI.