| Part one The diagnostic efficacy of radiomics model for predicting minimally invasive adenocarcinoma and invasive adenocarcinoma with mixed ground-glass nodulesObjective A radiomics model was constructed,and the data feature algorithm was used to quantify the phenotypic characteristics of the medical images of the tumor.Based on the radiomics model,the infiltration of lung adenocarcinoma with mixed ground-glass nodules in CT images was predicted and its application value was evaluated.Data and methods The clinical,pathological and imaging data of 120 patients who underwent chest CT examination and showed mGGNs(diameter ? 3 cm)in our hospital from January 2015 to June 2019 were retrospectively analyzed.All the patients were confirmed as lung adenocarcinoma by surgical resection.The images were imported into the lung nodule segmentation software(ITK-SNAP 3.8).The three-dimensional volume of interest(VOI)of the lesions were manually delineated for image segmentation,and the segmented images were imported into AK(Analysis Kit)software for image features.At the same time,25 cases were randomly selected.Two senior radiologist(A and B)segmented the nodules and extracted the radiomics features of the nodules,the consistency of lesion segmentation was assessed by evaluating the consistency of radiomics features extraction.The radiomics features were extracted by A twice and the consistency between the two times was evaluated.Then B extracted the radiomics features alone and evaluated the consistency between A(first time)and B extraction.The intraclass correlation coefficient(ICC)was used to evaluate the consistency of radiomics features extraction between A(twice)and A(first time)and B.The features are in good agreement if ICC is larger than 0.75.All cases were randomly divided into two groups.The ratio between the training group and the validation group was 7:3,with 85 nodules in the training group and 35 nodules in the validation group.AK software was used to extract radiomics features,including morphological features,histogram features,and texture features through the first-order,second-order and high-order statistical algorithms.The feature calculation and selection were performed in MatLab 2017b and R3.4.3 software.The dimensionality reduction process was performed by using a regular compression(Lasso)algorithm.After dimensionality reduction,the features with significant predictive value were retained,then the data was modeled.Rad score was calculated by prediction probability,which was classified as cut-off value.The radiomics model was established to evaluate the predictive diagnostic efficacy of the modelRESULTS A total of 120 mixed ground-glass nodules were enrolled,including 58 MIA and 62 IAC.The radiomics process includes image data acquisition,VOI segmentation,feature extraction,feature selection,and model establishment.In the 25 randomly selected cases,the ICC range of the radiologist A's two radiomics feature extraction was 0.84-0.95 and the ICC range of radiomics feature extraction between radiologist A(1st time)and radiologist B was 0.80-0.92,which indicated good consistency of radiomics features extraction within and between groups.A total of 396 radiomics features were extracted by the feature extraction method above.There are 20 morphological features,42 histogram features and 334 texture features.After the multi-step screening,the final dimensionality was reduced to 18 non-zero coefficient best diagnostic features,comprising ShortRunEmphasis_angle90_offsetl,ClusterShade_angle135_offset7,Correlation_AllDirec tion offset7_SD,HaralickCorrelation_angle45_offset7,GLCMEnergy_AllDirection_offset7,Sphericity,RunLengthNonuniformity_AllDirection offset4 SD,ClusterProminence_AllDi rection_offset7_SD,GLCMEnergy_angle45_offset7,RunLengthNonuniformity_AllDirectio n offsetl_SD,GLCMEntropy_AllDirection_offsetl_SD,GLCMEnergy_angle135_offset7,Correlation_angle135_offset4,GLCMEntropy_angle90_offset7,Inertia_angle45 offset7,Co rrelation AllDirection_offsetl_SD,histogramEnergy,ClusterProminence_angle45_offset7.These best diagnostic features were obtained by logistic regression analysis which had good efficacy in predicting MIA and I AC with mixed ground-glass nodules.The AUC for the training and verification groups were:AUC=0.87[95%CI:0.79-0.95]in the training group and AUC=0.85[95%CI:0.72-0.98]in the validation group.The sensitivity,specificity,accuracy,positive predictive value and negative predictive value of the training group were 0.854,0.818,0.835,0.814,and 0.857,respectively.The sensitivity,specificity,accuracy,positive predictive value,and negative predictive value of the group were 0.824,0.889,0.857,0.875,and 0.842,respectivelyConclusion The radiomics features constructed a unique model,which had high diagnostic efficiency for predicting the infiltration of lung adenocarcinoma,and high practical value for the identification of MIA and I AC.Part Two The diagnostic efficacy of CT feature model for predicting minimally invasive adenocarcinoma and invasive adenocarcinoma with mixed ground-glass nodulesOBJECTIVE By comparing the chest CT features of the mixed ground-glass nodules of minimally invasive adenocarcinoma and invasive adenocarcinoma.The differences of CT features between the them were explored to improve the accuracy of differential diagnosis.Materials and Methods 120 cases of mGGNs(including 58 cases of MIA and 62 cases of IAC)were the same as the first part.The morphological features of mGGNs were analyzed by two senior radiologist.The 3D-CT lung nodule segmentation software(FireVoxel software)was used to delineate the VOI and to measure the quantitative features of the nodules.The concordance correlation coefficient(CCC)was used for repeatability test.The quantitative features with high repeatability(coefficient value>0.9)were selected for analysis.The CT features included these parameters:age,gender,location,morphological features and quantitative features.Morphological features include shape,tumor boundary,lobulation,spiculation,pleural tag,vessel convergence sign,vacuole sign and bronchus sign.Quantitative features include the average nodule size,the average size of solid components,the ratio of the solid component and the average CT value nodules.Data processing was performed using SPSS 25.0 statistical software.The measurement data conformed to the normal distribution or the approximate normal distribution,and the results were expressed as mean±standard deviation.The comparison between the two groups was performed by independent sample t test;the count data was analyzed by chi-square test.Statistically significant(P<0.05)measurement data and count data were used for single-factor and multi-factor logistic regression analysis between groups.The independent predictors were selected to construct CT feature model.The ROC curves were drawn,and the optimal cut-off value,accuracy,sensitivity,and specificity were calculated.RESULTS Among 120 cases of mGGNs,the pathological results were 58 cases of MIA,accounting for 48.3%,and 62 cases of IAC,accounting for 51.7%.Comparison of morphological characteristics of mGGNs between the two groups:lesion shape(P=0.011),lobulation(P=0.001),spiculation(P=0.000),vacuole sign(P=0.013),bronchus sign(P=0.000),vessel convergence sign(P=0.000),the difference was statistically significant(P<0.05).Multivariate logistic regression analysis was performed on the statistically significant differences between the two groups.The difference in spiculation(P=0.013)was statistically significant.The quantitative characteristics of mGGNs in the two groups were compared.The average nodule size of MIA was 16.51±3.05mm.The average solid component size was 5.21 ±2.33mm.The ratio of solid components was 31.41±12.05%and the average CT value of nodules was-491±122.60HU.The average nodule size of IAC was 18.93±3.49mm.The average solid component size was 7.51±2.87mm.The ratio of solid components was 40.27±14.66%and the average CT value of nodules was-398.77±118.77HU.MIA and IAC were compared:the average nodule size(P=0.000),the average solid component size(P=0.000),the ratio of solid components(P=0.000)and the average CT value of nodules(P=0.000),which were statistically significant between the two groups.By multivariate logic analysis,the average nodule size(P=0.001),the average solid component size(P=0.015),the ratio of solid components(P=0.003)and the average CT value of nodules(P=0.003).They all had statistical significance.ROC curves were plotted for these statistically significant factors between the two groups.The average nodule size had an AUC of 0.695,an optimal cut-off value of 19.22 mm,an accuracy of 0.742,a sensitivity of 0.677 and a specificity of 0.799.The AUC of average solid component size was 0.743,the optimal cut-off value was 5.23mm,the accuracy was 0.708,the sensitivity was 0.758 and the specificity was 0.655.The AUC of solid component ratio was 0.695,the optimal cut-off value was 30.04%,the accuracy was 0.700,the sensitivity was 0.758 and the specificity was 0.655.The AUC of average CT nodule value was 0.735,the optimal cut-off value was-495.95,the accuracy was 0.712,the sensitivity was 0.742 and the specificity was 0.697.Conclusion The average nodule size,the average solid component size,the ratio of solid components,the average CT value of nodules and spiculation of mGGNs were helpful for the differential diagnosis of MIA and IAC.mGGNs with spiculation and large quantitative parameters were more prone to I AC.Part Three The diagnostic efficacy of Nomogram model for predicting minimally invasive adenocarcinoma and invasive adenocarcinoma with mixed ground-glass nodulesOBJECTIVE By combining Rad score with the CT features,a Nomogram model was constructed,which was a multi-factor logistic regression model for detecting Rad score and CT features.The radiomics nomogram was drawn and corresponding to each feature.The score was calculated and the risk factor was calculated.To compare the differences in diagnostic efficacy and clinical net benefit between the Nomogram model,the radiomics model and the CT feature model,and to explore the diagnostic value of the Nomogram model for predicting mixed ground-glass nodules with MIA and IAC.Materials and Methods 120 cases of mGGNs(including 58 cases of MIA and 62 cases of IAC)were the same as the first and second parts.The Rad score obtained in the first part and the CT features in the second part were analyzed by logistic regression analysis to obtain the combined Nomogram model.The Nomogram model was constructed in the training group,and the model was validated in the corresponding verification group.The training group and the verification group was at a ratio of 7:3.The radiomics nomogram was drawn.The consistency test was performed by plotting the calibration curve of the predicted value and the actual value in the training group and the verification group.The Hosmer-Lemeshow test was performed to observe whether the predicted model was significantly different from the real result.The ROC curves were used to evaluate the differential diagnosis of MIA and IAC with mixed ground-glass nodules between the Nomogram model,the radiomics model and the CT feature model.The sensitivity,specificity,accuracy,positive predictive value and negative predictive value of the two models were calculated.Based on the results of the analysis,the R3.4.3 software was used to establish the radiomics nomogram for predicting the risk of mGGNs.According to each risk factor,corresponding to the upper ruler(0?100),you could get the corresponding score of the factor.Decision curve analysis(DCA)was used to calculate the net benefit of the threshold probability range in both the training and verification groups,and to quantify the clinical net benefit with different threshold probabilities in the verification group.It compared with positive samples that were treated with intervention and negative samples that didn't receive intervention to assess the clinical value of the nomogram model.RESULTS The average nodule size,the ratio of solid components,the average CT value of nodules and Rad score of mGGNs could be used as independent predictors for predicting MIA and IAC with mixed ground-glass nodules.The AUC was 0.91[95%CI:0.85-0.97]for the training group and the AUC was 0.89[95%CI:0.77-1.0]for the verification group in CT feature model.The AUC was 0.93[95%CI:0.88-0.99]for the training group and the AUC was 0.99[95%CI:0.96-1.0]for the verification group in Nomogram model.The sensitivity,specificity,accuracy,positive predictive value and negative predictive value was 0.884,0.857,0.871,0.864,0.878 for the training group and was 0.739,0.812,0.800,0.844,0.747 for the verification group in CT feature model.The sensitivity,specificity,accuracy,positive predictive value and negative predictive value was 0.889,0.900,0.894,0.909,0.878 for the training group and was 0.944,0.941,0.943,0.944,0.941 for the verification group in Nomogram model.As a result,the AUC of the Nomogram model was superior than the radiomics model(in part one)and the CT feature model.The predicted models obtained from the calibration curves in the training group and the validation group were in good agreement with the observed results.The Hosmer-Lemeshow test showed that there was no significant difference between the prediction model and the real result(P>0.05).The net benefit of the model was superior to the CT feature model,the positive sample receiving intervention therapy,and the negative sample not receiving intervention therapy in the threshold range from 0.038 to 0.724.Conclusion For predicting MIA and IAC with mixed ground-glass nodules,the Nomogram model had superior performance compared with the radiomics model and the CT feature model.It showed better predictive performance and it's an effeective,practical differential diagnosis method. |
PDF Full Text Request