Computer-aided Diagnosis And Classification Study On The Liquid-based Cytology Of Cervical Intraepithelial Neoplasia

OBJECTIVETo reveal the quantitative morphological parameters and colorimetric parameters of cervical intraepithelial in liquid-based cytology, screened the morphological and colorimetric parameters which have important value of liquid-based cytology diagnosis. Explored the significance of these parameters in the diagnosis of cervical intraepithelial neoplasia and established a new method for quantitative diagnosis of cervical intraepithelial neoplasia according to the morphological and colorimetric parameters.METHODSLiquid-based cytology slices from 40 cases of normal people and 224 cases patients diagnosed with cervical intraepithelial neoplasia were collected from Huayin Medical Testing Center of Southern Medical University and Huizhou Central People’s Hospital. The collected slices were divided into 5 groups including negative for intraepithelial lesion or malignancey (NILM), atypical squamous cells of undetermined significance (ASC-US), low-grade squamous intraepithelial lesion (LSIL), and atypical squamous cells cannot exclude HSIL (ASC-H), high-grade squamous intraephithelial lesion (HSIL), the cells images were collected under the optical microscope at 40 times objective lens and input in computer. ImageJ 1.48 image analysis software was used to test the morphological and colorimetric parameters of the whole cell and nucleus, these selected morphological parameters including area, major axis, minor axis, perimeter, form factor PE, form factor AR, regular form factor, Form Irregular Index, Nuclear to Cytoplasmic Area Ratio and these selected colorimetric parameters including red, green, blue, red color coefficient, green color coefficient, blue color coefficient.Statistical analysis of morphological parameters and colorimetric parameters of cervical liquid-based cells in different groups were performed by SPSS 17.0 software. Results of continuous variables were shown as mean (), standard deviation (SD), maximum (Max), minimum (Min), full range (Range) and coefficient of variation (CV). Test results of different groups were analyzed as follows:One-Way ANOVA analysis was used to compare the differences among different groups to meet the test conditions of parameters, variables that do not meet the test conditions were compared using the U Mann-Whitney test or the Kruskal-Wallis test. The stepwise discriminant analysis was conducted among different groups, and the discriminant functions and Fisher linear discriminant functions were modelling respectively, the classification map were drawn and the discrimination accuracy as well as validation accuracy of each group were calculated. Logistic regression analysis and ROC curves were performed for two specific groups to screen the parameters that had significant statistical difference. Calculated the sensitivity and specificity of classification and cut point when obtained the maximum Youden’s Index. According to the parameter values obtained from the calculation, the inclusion criteria were re-adjusted and the stepwise discriminant functions were re-modelling between different groups, internal-validation accuracy and cross-validation accuracy of each groups were calculated.RESULTS1. The test results of morphological and colorimetric parameters of cervical liquid-based cytologyIn the comparison of morphological parameters, NILM cell area was the largest and its mean nucleus area was the smallest among all groups. The difference was statistically significant (P<0.001). The coefficient of variation of cell area and nucleus area of ASC-H and HSIL were much higher than that of the other three groups. The difference was statistically significant (P<0.001). The average size of ASC-US and LSIL cells were nearly the same size while the average nucleus area of the former was larger than the latter, and the coefficient of variation was less than that of the latter. The coefficient of variation of each parameter of the NILM cells and nucleus were the smallest among all groups. The difference was statistically significant (P<0.001). The Form Irregular Index of HSIL and ASC-H was the largest in the different groups. The difference was statistically significant (P<0.001). The coefficient of variation of Form Factor AR of ASC-US and coefficient of variation of Form Factor PE of HSIL were both the largest in each group. The difference was statistically significant (P<0.001). The coefficient of variation of nucleus and Nuclear to Cytoplasmic Area Ratio of HSIL and ASC-H were much higher than the other three groups. The difference was statistically significant (P<0.001).In comparison of colorimetric parameters, the nucleus’s red, green and blue color in the average value of NILM the largest in each group. The difference was statistically significant (P<0.001). The coefficients of variation of red and blue colors in each group were the smallest. The difference was statistically significant (P<0.001). The red, green and blue colors in the average value of HSIL and ASC-H in each group were the smallest. The difference was statistically significant (P<0.001).2. Discriminant analysis based on morphological and colorimetric parameters of cervical intraepithelial neoplasia(1) Discriminant analysis of normal squamous epithelial cells and lesions squamous cellThe internal-validation and cross-validation accuracy of normal squamous epithelial cells and lesions squamous epithelial cells were both 96.5%, and the stepwise discriminant functions were established by morphological and colorimetric parameters. Discriminant functions of normal squamous epithelial cells of internal-validation and cross-validation accuracy were 97.2% and 97.1% respectively. The discriminant functions were 96.4% and 96.3% for the internal-validation and cross-validation accuracy of the lesion squamous cells.(2) Discriminant analysis of different types of cervical intraepithelial neoplasiaThe overall internal-validation and cross-validation accuracy of different types of cervical intraepithelial neoplasia liquid-based cells was 74.9% and 74.6% respectively, the discriminant functions were established by morphological and colorimetric parameters. The internal-validation and cross-validation accuracy of NILM cells were both 92.7%, the internal-validation and cross-validation accuracy of ASC-US were 71.0% and 70.7% respectively, the internal-validation and cross-validation accuracy of LSIL were 71.3% and 70.8% respectively, the internal-validation and cross-validation accuracy of ASC-H were 72.8% and 72.4% respectively, the internal-validation and cross-validation accuracy of HSIL were 69.3% and 68.7% respectively.(3) Analysis and results of the two classification Logistic regression analysis and ROC curveBy two Logistic regression analysis results shown that Form Factor AR of cell, Form Irregular Index of nucleus, blue color had better identification effect between ASC-US and LSIL, the total accuracy of the regression equation was 77.3%. Two Logistic regression analysis results shown that area of cell, perimeter of cell, major axis of nucleus, minor axis of nucleus and green primary coefficient had better identification effect between ASC-H and HSIL, the total accuracy of the regression equation was 74.6%.The results of screening by ROC curve shown that Form Factor PE of cell, Form Factor AR of cell, Regular Form Factor of cell, Form Factor PE of nucleus, Form Factor AR of nucleus and blue primary coefficient had better identification effect between ASC-US and LSIL, when the blue primary coefficient was 0.425 could be achieved the maximum Youden’s index in all parameter (25.83%). The results of screening by ROC curve shown that area of cell, major axis of cell, minor axis of cell, perimeter of cell, area of nucleus, major axis of nucleus, minor axis of nucleus, perimeter of nucleus and blue color had better identification effect between ASC-H and HSIL, when blue color was 103.3 could be achieved the maximum Youden’s index while area of cell was 85.71 μm2 could achieved the second largest Youden’s index.(4) Discriminant functions were established with the sampling inclusion criteria was blue primary coefficient=0.425According to the two classifications Logistic regression analysis and ROC curve results shown that blue primary coefficient could achieve the best identification effect between ASC-US and LSIL. The ASC-US samples that blue primary coefficient was greater than 0.425 were removed and the LSIL samples that blue primary coefficient was less than 0.425 were removed, and then established stepwise discriminant functions between different groups. The results of the discriminant accuracy shown that the accuracy rate of ASC-US and LSIL was improved, the internal-validation and cross-validation accuracy of ASC-US were 88.7% and 88.5% respectively, the internal-validation and cross-validation accuracy of LSIL were 82.4% and 82.0% respectively, the overall internal-validation and cross-validation accuracy was 80.5% and 80.1% respectively.(5) Discriminant functions were established with the sampling inclusion criteria was cell area=85.71 μm2According to the two classifications Logistic regression analysis and ROC curve results shown that cell area could achieve the best identification effect between ASC-H and HSIL. The ASC-H samples that cell area was greater than 85.71 μm2 were removed and the HSIL samples that cell area was less than 85.71 μm2 were removed, and then established stepwise discriminant functions between different groups. The results of the discriminant accuracy shown that the accuracy rate of ASC-H and HSIL was improved, the internal-validation and cross-validation accuracy of ASC-H were 96.9% and 96.6% respectively, the internal-validation and cross-validation accuracy of HSIL were 87.4% and 87.0% respectively, the overall internal-validation and cross-validation accuracy was 82.9% and 82.5% respectively.(6) Discriminant functions were established with the sampling inclusion criteria was blue primary coefficient=0.425 and cell area=85.71μm2Re-established the stepwise discriminant functions with the sampling criteria were blue primary coefficient=0.425 and cell area=85.71μm2. The results shown that the internal-validation and cross-validation accuracy of each groups were improved. The internal-validation and cross-validation accuracy of NILM were 95.1% and 95.0% respectively, the internal-validation and cross-validation accuracy of ASC-US were both 90.5%, the internal-validation and cross-validation accuracy of LSIL were 84.4% and 84.1% respectively, the internal-validation and cross-validation accuracy of ASC-H were 97.2% and 96.9% respectively, the internal-validation and cross-validation accuracy of HSIL were 88.3% and 88.0% respectively, the overall internal-validation and cross-validation accuracy were 91.2% and 91.0% respectively.CONCLUSIONS1. There was a significant difference in the morphological and colorimetric parameters between different types of cervical intraepithelial neoplasia, and could be used for the computer quantitative analysis of cervical intraepithelial neoplasia.2. Stepwise discriminant functions had a better effect on the classification of normal squamous epithelial cells.3. The blue primary coefficient, cell area could be used for differential diagnosis between ASC-US and LSIL, ASC-H and HSIL.4. The discriminant accuracy of stepwise discriminant function among different types of cervical intraepithelial neoplasia cells could be improve after the sampling criteria were adjusted.