PI-RADS Score And Radiomic-Based Models For Detection Of Prostate Cancer

BackgroundThe PI-RADS score is of great significance in the diagnosis and evaluation of PCa as a traditional imaging mode.Radiomic is an emerging research method,and its machine learning model can help identify benign and malignant prostate lesions and provide invasive information and predict biochemical recurrence.Objectives1.This meta-analysis was undertaken to review the diagnostic accuracy of PI-RADS V2 for PCa detection with mp-MRI.2.To determine the inter-reader consistency and diagnostic performance of V2.1 in differentiating PCa using mp MRI.3.To investigate the diagnostic value of PI-RADS V2.1 for PCa at different serum PSA levels.4.To develop and validate two radiomics-based machine learning models(R Tree and Logistic Regression)for differentiating Cs PCa from Ci PCa compared with PI-RADS V2.1 score.Methods1.A comprehensive literature search of electronic databases was performed by two observers independently.The databases includes: PUBMED,EMBASE,Web of Science,Cochrane Library and Chinese database(CNKI,Wangfang,CBM,VIP).Inclusion criteria were original research using the PI-RADS V2 system in reporting prostate MRI.The methodological quality was assessed using the QUADAS-2 tool.Data necessary to complete 2 × 2 contingency tables were obtained from the included studies.The statistical computations were performed using STATA 12.0 software to pool sensitivity,specificity,DOR,LR+ and LR-with 95% CIs.If there was obvious heterogeneity,we found the sources of heterogeneity by subgroup analysis.2.In all,442 consecutive patients underwent mp MRI and subsequent systematic plus targeted biopsies were included.Clinical parameters and the score of PI-RADS V2.1 based mp MRI were investigated.Two readers independently analyzed the images with Version 2.1.The inter-reader agreement was calculated using Kappa statistics,and the diagnostic performance of Version 2.1 was analyzed by ROC curve,accuracy,sensitivity and specificity.If associations are detected in univariate analysis,multivariate logistic regression was also be used.3.442 patients who underwent prostate mp MRI and obtained pathological results were included in the study.The study population were divided into 3 groups according to the serum PSA level.The patients with PSA 4~10ng/ml were in group A,10~20ng/ml in group B and >20ng/ml in group C.Two readers performed V2.1 score on MRI images of patients in each group.The consistency of the results was evaluated by using Kappa statistic.The Receiver Operating Characteristic curve was drawn to evaluate the diagnostic efficacy of V2.1 for PCa.The AUC,sensitivity,specificity were calculated.4.In all,142 patients with pathologically confirmed PCa were enrolled,including 101 patients with Cs PCa and 41 patients with Ci PCa.All patients underwent T2 WI,DWI,and DCE before biopsy.The PI-RADS V2.1 score was used to evaluate mp-MRI images of 142 PCa patients by two readers with different diagnostic experience in prostate MRI diagnostic imaging.A Kappa test was used to assess the agreement between the two readers.The lesion was segmented using ITK‐SNAP software,and the three-dimensional VOI of the lesion was manually delineated by one doctor.A.K software was used for feature extraction on the obtained VOIs.A total of 402 radiomics features(including histograms,texture features,etc.)in 6 categories were automatically extracted from the ADC and T2 WI VOIs of each patient.The Spearman correlation coefficient and LASSO regression analysis were used to reduce the feature dimension and filter.And then the two mainstream machine learning algorithms(R Tree and Logistic regression analysis)were used to construct models for the selected feature parameters.Based on four different feature sets:(1)ADC feature set,(2)T2WI feature set,(3)T2WI and ADC joint feature set [screening and then combining,ie(1)+(2)],(4)T2WI+ADC feature set [The T2 WI and ADC sequences combined with the 804 radiomics features to select the optimal feature set(first combination and then screening)].Use training data to build R Tree and Logistic regression analysis prediction models to verify the validation set,and apply ROC to evaluate individual sequence diagnostic models,joint diagnostic models,and PI-RADS V2.1,and calculate the AUC,accuracy,specificity,and sensitivity.The Delong-test is used to compare the diagnostic effectiveness between different algorithms and models.P<0.05 indicates that the difference is statistically significant.Results1.Thirteen studies(2,049 patients)were analyzed.This is an initial meta-analysis of PI-RADs V2 and the overall diagnostic accuracy in diagnosing PCa was as follows: pooled sensitivity,0.85(0.78–0.91);pooled specificity,0.71(0.60–0.80);LR+,2.92(2.09–4.09);pooled LR–,0.21(0.14–0.31);pooled DOR,14.08(7.93– 25.01),respectively.Positive predictive values ranged from 0.54 to 0.97 and negative predictive values ranged from 0.26 to 0.92.Subgroup analysis showed that PI-RADS 3 scores as cut-off [0.89(0.81-0.94)/0.71(0.55-0.84)] has higher pooled sensitivity and specificity than with 4 score as cut-off value [0.81(0.74-0.86)/0.66(0.48-0.80)].2.Patients were histological proven in 245 cancers and 197 benign lesions.There is significant correlation between higher Version 2.1 score and the presence of PCa(P<0.05).For Cs PCa,with score≥4 as a threshold,accuracy,sensitivity and specificity was 0.848,0.943,0.777 for reader 1 and0.835,0.922,0.766 for reader 2,respectively.The diagnostic concordance for PCa and Cs PCa was good(Kappa=0.732-728).Inter-reader agreement for TZ(Kappa=0.633)was lower than for PZ(Kappa =0.768).Multivariate analysis revealed that V2.1 score was the only significant parameter for both readers [odds ratio: 36.56(15.99-107.13);31.88(14.24-91.10),P <0.001,respectively].3.There were significant differences in age and PCa prevalence among the three groups(P <0.001).Kappa values improved with increasing PSA levels(Kappa values were: 0.547,0.709,0.763,respectively).The AUC value improved with increasing PSA levels(AUC values were: 0.835,0.873,and 0.921,respectively).The sensitivity of V2.1 to the detection of PCa in each group were 0.926,0.9,0.95,and the specificity were 0.65,0.71,0.64,respectively.4.ModelT2WI(constructed from 7 optimal T2 WI feature sets),Model ADC(constructed from 4 optimal ADC feature sets),Model T2WI/ADC(constructed from a total of 4 feature sets selected by the two sequences)and Model T2WI+ADC(constructed from 4 optimal feature sets selected from 804 features combined by two sequences).Comparison of the diagnostic performance of the four models obtained by applying the two algorithms: The results show that the differential diagnostic model established by the Logistic regression analysis method has better performance than the R Tree and the difference is statistically significant(P<0.05).The AUC values,accuracy,sensitivity,specificity of the four models of the training set obtained by the radiomics Logistic regression are 0.847,0.84,0.903,0.692,respectively;0.863,0.88,0.967,0.69,respectively;0.761,0.7,0.838,0.384,respectively;0.868,0.795,0.806,0.769,respectively.The AUC values,accuracy,sensitivity,specificity of the four models of the test set are 0.908,0.84,0.935,0.615,respectively;0.863,0.886,0.967,0.692,respectively;0.762,0.704,0.838,0.385,respectively;0.868,0.795,0.806,0.77,respectively.Among them,the highest AUC value is the Model ADC.The lowest AUC value is the combination of two sequences Model T2WI/ADC.Taking PI-RADS V2.1 of 5 score as the cut-off value,the AUC values,accuracy,sensitivity and specificity of the two readers diagnosis performance were 0.767,0.725,0.703,0.78,respectively and 0.813,0.76,0.713,0.878,respectively.Conclusion1.Currently available evidence indicates that PI-RADS V2 appears to have good diagnostic accuracy in patients with PCa lesions with high sensitivity and moderate specificity.However,no recommendation regarding the best threshold can be provided because of heterogeneity.2.PI-RADS V2.1 appears to have good diagnostic performance in Cs PCa with category 4 as the threshold and it was the only independent predictor of Cs PCa.Both readers have good inter-reader reliability.3.PI-RADS V2.1 showed excellent performance in predicting PCa in every PSA zone.With the increase of PSA,its diagnostic efficacy also improves.4.Both the T2 WI and ADC based radiomics models showed high diagnostic efficacy and outperformed the PI-RADS V2.1 score in distinguishing Cs PCa and Ci PCa.Especially the machine learning algorithm Logistic model performs better.