Application Of Bi-parametric MRI Radiomics In PI-RADS 3 Lesions

Part I Bi-parametric MRI radiomics predicting the malignancy of PI-RADS 3 category lesions Objective:To explore the value of bi-parametric magnetic resonance imaging(MRI)radiomics in predicting malignancy of prostate imaging reporting and data system(PI-RADS)3 category lesions by constructing and comparing differents models.Methods:The clinical and imaging data of 232 patients with PI-RADS score of 3 from the First Hospital of Shanxi Medical University between January 2018 and October 2021 were retrospectively analyzed.All patients had clear pathological results after prostate biopsy or radical prostatectomy and were divided into training set(n=162)and validation set(n=70)randomly at a ratio of 7:3.The region of interest(ROI)was delineated layer by layer on T2 WI,DWI and ADC sequences manually by two radiologists using ITK-snap software,and the Radiomics module in 3D slicer software was used to extract radiomics features.The stability of features was estimated by the intra-and inter-group correlation coefficients(ICCs)and the ICC ≥ 0.75 was considered to be consistent.The progress of radiomics feature selection was performed by univariate logistic regression and least absolute convergence and selection operator(LASSO)regression,tenfold cross validation was used to adjust the parameter λ to determine the number of features,and univariate analysis was used to obtain clinical independent risk factors that could predict the outcome.The clinical model,radiomics model and radiomics score(radscore)were calculated by multivariate logistic regression.Further,a clinical-radiomics combined model was constructed by combining the radscore with clinical risk factors.The diagnostic efficacy of the three models was analyzed and compared with the area under the receiver operating characteristic(ROC)curve and its derived indicator area under the curve(AUC),sensitivity,specificity and accuracy.The added clinical benefits between the models was quantified with net reclassification improvement(NRI).Finally,select the best performing model as the prediction model,plot the calibration curve and decision curve to assess the correction and clinical practicability of the model,and plot the nomogram to realize the model visualization.Results:Among 232 patients with PI-RADS 3 lesions,there were 62 cases of prostate cancer(PCa)with age of 67.71 ± 7.37 years and 170 cases of benign prostate disease with age of70.08 ± 7.67 years.There was no significant difference between the training set and the test set for each clinical indication(p>0.05).A total of 2553 radiomics features were extracted from each patient.After ICC processing,the remaining 2164 stable features were used for subsequent analysis.Univariate logistic regression and LASSO regression showed that 2 of T2 WI features,4 of DWI features and 7 of ADC features were related to the malignancy of PI-RADS 3 lesions.Age,prostate volume(PV)and prostate specific antigen density(PSAD)were selected as clinical independent factors.Among the three models,the combined model binding clinical and radiomics features achieved the best prediction performance.The AUC,sensitivity,specifiity and accuracy in the training set and validation set were 0.924,0.822,0.880,0.864 and 0.946,0.941,0.925 and 0.929 respectively.The NRI of the combined model in training set and validation set were 0.582,0.425 and 0.792,0.523,respectively,compared with the clinical model and the radiomics model.Calibration curve and decision curve showed that the model correction(Hosmer-Lemeshow test: training set P=0.784,validation set P =0.107)and clinical practicability was good.Conclusion:The bi-parametric MRI radiomics has certain predictive value for the malignancy of PI-RADS 3 category lesions,and the clinical-radiomics model based on age,PV,PSAD and radscore is expected to recognize PCa of PI-RADS 3 lesions and assist clinical decision-making.Part II A clinical-radiomics fusion model predicting the aggressiveness of PI-RADS 3 lesionsObjective: To establish and validate a fusion model based on clinical and radiomics features to further predict the aggressiveness of PCa in prostate imaging report and data system(PI-RADS)3 lesions.Methods: The clinical,imaging and pathological information of 62 cases with malignant PIRADS 3 lesions in the first part of the study were analyzed retrospectively.According to Gleason score,the lesions were classified into clinically significant prostate cancer(cs PCa,GS ≥ 3+4,highly aggressive)group and clinically insignificant prostate cancer(ci PCa,GS=3+3,low aggressive)group.The training set(n=41)and validation set(n=21)were split by the ratio of 2:1.The clinical risk factors were determined by Akaike information criterion(AIC)through backward step regression.2164 stable radiomics features processed by ICC in the first part were used.The reducution of the features correlation and quantity dimension and the calculation of radiomics score(radscore)were carried out by pearson correlation analysis and recursive feature elimination(RFE).Finally,the clinical features and radscore were integrated to construct a clinical-radiomics fusion model using multivariate logistic regression.The predictive effectiveness of the fusion model was validated through ROC curve,calibration curve and decision curve analysis,and used nomogram to achieve the model visualization.Results: Among 62 patients with PCa of PI-RADS 3 lesions,there were 38 cases of cs PCa and 24 cases of ci PCa.Age was selected as a clinical independent risk factor.Pearson correlation analysis and RFE screened 4 of T2 WI features,1 of DWI features and 1 of ADC features.The radscore of cs PCa group was higher than that of ci PCa group(1.52 ± 2.649 and-1.815 ± 2.065,p<0.001).The AUC,sensitivity,specificity and accuracy of the clinical-radiomics fusion model in the training set and validation set were 0.879,0.736,0.870,0.838 and 0.908,0.944,0.811,0.868,respectively.The calibration curve showed that the model had good fit goodness in training set and validation set and the p-values of Hosmer-Lemeshow test were 0.907 and 0.689 respectively.The decision curve showed that the fusion model had higher clinical net benefit than that of all or no treatment schemes with the threshold value > 0.17.Conclusion: The clinical-radiomics fusion model based on age and radiomics features has certain predictive value for the aggressiveness of PI-RADS 3 lesions,which is conducive to identifying cs PCa and guiding patients to personalized treatment.