| Objective: The machine learning machine of radiosomics was trained using the texture features of mp MRI images,and the value of the diagnosis and identification of Gleason scores for prostate cancer was explored through the performance evaluation of the machine.Materials and Methods: A retrospective study of patients with prostate cancer who underwent mp MRI and confirmed pathologically and obtained a Gleason score from January 1,2015 to September 1,2018.The final study population was 219 patients.The patients’ T2 WI and DWI image data were exported from the PACS system,and the DWI image was processed using GE Functool software to obtain an ADC map.The pathological Gleason Score(GS)of each patient was confirmed and reviewed,matched by MRI and histopathological findings,and the approximate location of PCa was determined on the T2 WI image and ADC map.On the ITK-SNAP software,the region of interest(ROI)is manually layered along the lesion boundary to be finally fused to obtain three-dimensional data,the volume of interest(VOI).Using the One Key module in the AK software developed based on the MITK platform,396 texture parameter features were extracted.Divide the data into two groups,including GS(3+3)(n=36)vs GS(>7)(n=103)and GS(3+4)(n=43)vs GS(4+3)(n=49),each group were randomly divided into training and test group according to the ratio of 7:3.Then the redundant features are removed by the feature selection step.Finally,in the GS(3+3)vs GS(>7)groups,7 and 3 features are extracted based on the ADC map and T2 WI images,respectively,while in the GS(3+4)vs GS(4+3)group,three features were extracted based on both ADC and T2 WI images.Six machine learning models are established by selecting the features: SVM: C_SVC & LINEAR,SVM: C_SVC & RBF SVM: C_SVC& sigmoid,Logistic Regression,K Nearest Neighbor(KNN),Bayesian(Bayes).At the same time,these model classifiers are trained to obtain the accuracy,specificity and sensitivity of the classifier.The ROC curve is obtained by SPSS 25.0 using the texture parameter values after feature selection,and area under the ROC curve(AUC)and 95%confidence interval.Results: The optimal classifier was automatically selected from the six learning machines using the One Key module in the AK software,and the classification performance was evaluated using AUC.AUC values range from a minimum of 0.590 to a maximum of 0.915.The Logistic Regression classifier trained using the ADC map texture feature has an AUC value of 0.915,which shows excellent classification performance in the identification of GS(3+3)and GS(>7)groups;and in GS(3+4)vs GS(4+3)group,the AUC value of the C_SVC&RBF classifier based on the ADC map is0.661,and the classification performance is better compared with the AUC value based on the T90 WI image model of 0.590.In the same classifier trained using ADC map and T2 WI image texture feature parameters,the AUC values of learning machine based on the ADC map are higher than those using T2 WI image texture training,especially in GS(3+3)vs GS(> 7),the AUC value of the Logistic Regression machine learning model based on the ADC map is as high as 0.915,showing excellent classification performance.Conclusion: Establishing a machine learning machine based on radiomics features of multiparametric magnetic resonance imaging(mp MRI),then achieve a good classification performance for Gleason Score of prostate cancer.This has significant potential and clinical value for different treatments and prognosis predictions for prostate cancer patients,and it will likely become a diagnostic method for clinical and radiologists in the near future. |
