Establishment And Application Of A Prediction Model For Clinically Significant Prostate Cancer Based On Faster Region-based Convolutional Neural Network And Multiparametric MRI

Purpose:Based on the structure of Faster RCNN and mp MRI for prostate,we intend to construct,apply and evaluate a prediction model for csPCa,in order to improve the ability of mp MRI to predict csPCa.Method:Patients who underwent transperineal targeting + systematic puncture for biopsy of prostate in our hospital from May 2019 to May 2021 were included.A total of 341 patients were selected,excluding patients with incomplete data,poor image quality and previous related treatment.Among them,263 patients from May 2019 to December 2020 constitute the cohort 1,and 78 patients from January 2021 to May 2021 constitute the cohort 2.Cohort 1 included 84 patients with csPCa and 179 patients without csPCa.A total of 4763 MRI images including T2 WI,DWI(b=1500)and ADC were obtained from cohort 1,including 1551 with csPCa and 3212 without csPCa.Cohort 2 included 20 patients with csPCa and 58 patients without csPCa.A total of 2582 MRI images were obtained from cohort 2,including 375 with csPCa and 2207 without csPCa.Cohort 1 was randomly divided into training-validation set(80%)and internal test set(20%),and cohort 2 was used as external test set.All images were operated for format conversion,desensitization,registration and ROI tagging.We outlined and labelled the prostate area in each image according to the pathological examination.A Faster RCNN model composed of feature extraction network,region proposal network and ROI network was constructed.In the training and validation,the parameters are adjusted to obtain the final model with high accuracy and low value of loss function.The sensitivity,specificity,positive predictive value,negative predictive value,accuracy,ROC curve and AUC were obtained in two test sets to evaluate the ability of the model to predict csPCa.The prediction ability of the model was compared and integrated with the prediction ability of radiologists and other common clinical features.The statistical data were analyzed and plotted by python,SPSS,medcalc and R software.P < 0.05 means statistically significant.Result:During the training,the value of loss function decreases and stabilizes to 0.006,and the accuracy increases and stabilizes to 0.999.The model shows the accurate positioning and classification of csPCa in the test set.The AUC of the model in the internal test set is0.962;The AUC of the model in the external test set is 0.827.After grouping according to the sequence,the AUC of the model in single sequence is0.949(T2WI),0.969(DWI)and 0.996(ADC)in the internal test set;the AUC of the model in single sequence is 0.664(T2WI),0.720(DWI)and 0.797(ADC)in the external test set.In the external test set,when the csPCa model was used for prediction,compared with the prediction of radiologists,the AUC of the whole,DWI and ADC increased by 0.035,0.004 and 0.016 respectively,and the AUC of T2 WI decreased by 0.088.When the radiologists combined with the model predicted,compared with the prediction of radiologists,the AUC of the whole,T2 WI,DWI and ADC increased by 0.055,0.087,0.005 and 0.035 respectively.There was no statistically significant difference.In univariate analysis,the AUC of the model was higher than that of TPSA(0.827 vs0.614,P=0.013),PV(0.827 vs 0.694,P=0.076),and lower than that of PI-RADS(0.827 vs 0.846,P=0.778).The statistically significant clinical features of univariate analysis were included in multivariate analysis.Finally,PI-RADS and scores of the model were independently correlated with the existence of csPCa.Based on PI-RADS and score of the model,a comprehensive model is constructed.The decision curve shows that the comprehensive model has the highest net benefits when the risk threshold is less than 0.8in most cases.The AUC of comprehensive model was higher than that of TPSA(0.901 vs0.614,P < 0.001),PV(0.901 vs 0.694,P=0.003),PI-RADS(0.901 vs 0.846,P=0.058)and the model(0.901 vs 0.827,P=0.068).Conclusion:We build prediction model for csPCa based on Faster RCNN and mp MRI,which can complete the prediction task for csPCa.The model has good prediction performance in DWI and ADC sequences,but its performance in T2 WI sequences needs to be improved.The prediction ability of the model is similar to that of the radiologists.With the help of the model,the prediction ability of the radiologists is improved in each sequence,which is not statistically significant.The score of the model and PI-RADS can be used as independent predictors of csPCa.The combination of the two can achieve higher net benefits of clinical decision-making.The model has higher prediction ability than common clinical features(TPSA).The comprehensive model has higher prediction ability than common clinical features(TPSA,PV).