Identification And Risk Stratification Of Prostate Cancer By MRI Radiomics Based On T₂WI And DWI Sequences

Objective:To establish a radiomics model based on T₂WI and DWI sequences to differentiate benign and malignant prostate lesions and perform risk stratification,and to explore the diagnostic efficiency and value of different machine learning models.Materials and Methods:A total of 271 patients with prostate lesions confirmed by pathology in our hospital were retrospectively collected,of which 133 were benign and138 were malignant.In the malignant group,83 patients had GG>2 and 55 patients had GG≤2.All patients underwent preoperative MRI examination.ITK-SNAP software was used for 3D delineation,and the urethra,hemorrhage and calcification should be avoided when drawing the ROI.The same delineation criteria was used for all lesions to ensure consistent areas of interest in T₂WI and ADC images.If a patient has multiple lesions with different Gleason scores,the lesion level with the highest Gleason score and the largest diameter is selected to obtain 3D data by manually outlining the area of interest.The radiomics features outlined in VOI were extracted using the Pyradiomics package in Python.To ensure the uniform feature scale,all radiomics features were standardized by z-score normalization,and the features were selected by the maximum correlation-minimum redundancy（m RMR）algorithm.The redundant and irrelevant features were narrowed by m RMR,and the optimized feature subset was selected to construct the final radiomics model.The scope of redundant and irrelevant features was narrowed by m RMR,and the optimized feature subset was selected to construct the final radiomics model.R software was used to complete the statistical analysis of radiomics.The data were randomly divided into the training group and the test group according to 7:3,and the random forest（RF）,support vector machine（SVM）,extreme gradient boosting（XGboost）and generalized linear model（GLM）machine learning algorithms were used to construct a model for the above screened characteristic parameters.Three sets of models were constructed:1)ADC feature set;2)T₂WI feature set;3)T₂WI+ADC feature set.The hyperparameters of the above machine learning algorithms are adjusted by 10-fold cross-validation and heuristic algorithms.ROC curve was used to evaluate the effectiveness of different models in test data sets.Delong’s test to compare the AUC difference of ROC curve。P<0.05 was considered statistically significant.Result1.Feature extraction and screeningBased on ADC and T₂WI images,1317 features were extracted respectively.The m RMR algorithm was used to screen the acquired image omics features.In the benign and malignant group,20 features were screened based on ADC,20 features were screened based on T₂WI,and 20 features were screened based on T₂WI+ADC.In the differentiation between GG≤2 and GG>2 PCa group,20 features were selected based on ADC,18 features were selected based on T₂WI,and 18 features were selected based on T₂WI+ADC.2.Model construction and effectiveness evaluationAccording to the selected radiomics features,the radiomics models were established based on RF,SVM,XG Boost and GLM classifiers.In distinguishing benign and malignant lesions,SVM model with ADC feature set had the highest diagnostic efficiency,with AUC value and accuracy of 0.87 and 0.79,respectively.The diagnostic efficiency of XGboost model in the feature set of T₂WI+ADC was the highest in PCa with GG≤2 and GG>2,and the AUC value and accuracy were 0.86and 0.83,respectively.De Long’s test showed that there were no significant differences between the benign and malignant lesion groups and the intra-group models of GG≤2and GG>2 groups（P>0.05）.Conclusion1.The machine learning model based on radiomics can effectively differentiate benign and malignant prostate lesions and risk stratification.2.SVM was the best classification model for the group of benign and malignant prostate lesions.3.XGboost was the best classification model for the GG≤2 and GG>2 PCa group,and SVM also achieved good results.