The Construction Of Radiologic Extranodal Extension Radiomics And Deep Learning Model And Its Prediction Value For The Aggressiveness Of Prostate Cancer

Part 1 Correlation study between rENE and pathological grading of prostate cancerObjective: To explore the association between radiological extranodal extension（rENE）and pathological grades in prostate cancer（PCa）patients and to investigate whether rENE could predict pathological results independently.Methods: We retrospectively identified patients who underwent pelvic MRI and pathological diagnosed PCa disease,RP and PLND diagnosed N1 staging（p N1）patients or MRI diagnosed N1 staging（r N1）patients from January 2004 to December 2021 in our center,a total of 2143 patients were included.According to the revised criteria for HNSCC,1505 patients were diagnosed as rENE+ and 638 patients were diagnosed as rENE-.The significant differences were analyzed by the Wilcoxon rank-sum test.The correlation indices between rENE and the GS and ISUP grades were calculated.Logistic regression was used to determine if rENE was an independent predictor of the ISUP≥4 group.Results: 1.There was a good inter-reader agreement between r N1 and rENE sign interpretation and diagnosis.2.GS and ISUP grades in rENE+ groups were significantly higher than rENE-groups in p N1,r N1 and all cases（P<0.01）.3.By Spearman’s rank correlation analysis,GS and ISUP grades were positively correlated with rENE in p N1,r N1,and in all cases,the correlation coefficients were 0.57,0.53,0.56,0.53 and 0.45,0.41,respectively.4.Univariate and multivariate regression analyses of all cases,p N1 and r N1 groups suggested that rENE,T-stage,and t PSA were independent predictors of the ISUP≥4 group,and rENE had the highest predictive power.Conclusion: rENE is positively related to GS and ISUP grades in PCa patients.GS and ISUP grades in the rENE+ group were significantly higher than rENE-group.rENE could be an independent predictor for ISUP≥4 group patients.Part 2 Construction of the rENE radiomics model and its prediction of prostate cancer aggressivenessObjective: To construct the radiomics model of rENE,and evaluate its value in predicting the aggressiveness of PCa.Methods: From January 2017 to June 2022,160 patients diagnosed with PCa N1 staging were collected retrospectively in the Xijing Hospital,Air Force Medical University.110 cases were classified as the rENE+ group and 50 cases were classified as rENE-group according to the strict standard of rENE in the newest AJCC TNM staging of HNSCC.They were further divided into a training set（112 cases,77 cases in the rENE+ group and 35 cases in rENE-group）and a testing set（48 cases,33 cases in the rENE+ group,and 15 cases in rENE-group）by a random number table method.The differences in baseline data between groups were tested by the Wilcoxon rank sum test and Mann-Whitney U test.3D slicer software was used to segment images,and FEA software was used to extract the radiomic features（RF）and establish the radiomic model.Analysis of variance（ANOVA）and F value were used to select and sort the orders of RF.The top 20 weighted RF were selected and the models were built by 2 means of normalize（Z-score and Mean method）and 8 classifiers（SVM,LDA,LR,Adaboost,AE,RF,Lasso regression,and NB）.5-fold cross validation set was constructed to validate the models.The ROC curve and the AUC were used to evaluate the diagnosing value of the models.Combined AUC values of CV validation set and testing set with model stability to screening for the most optimal model.ROC curve was used to evaluate the value of the model in predicting ISUP≥4 grading groups.The logistic regression was used to evaluate the independent predictor of the ISUP≥4 grading group.Result: 1.The final radiomic model of rENE was Zscore_PCC_ANOVA_2NB,which was constructed by 2 RF.The AUCs of diagnosing rENE were 0.954,0.952,and 0.960 respectively in the training group,testing group,and cross-validation group.2.The AUC of the Zscore_PCC_ANOVA_2NB model in predicting ISUP grades ≥4 in the training group,testing group,and the whole group were 0.862,0.944,and 0.923 respectively.3.The independent factors of the ISUP≥4 group were Zscore_PCC_ANOVA_2NB model and t PSA by multiple logistic regression.Conclusion: The radiomic model of rENE we constructed could reduce inter-observer differences effectively and could predict the aggressiveness of PCa noninvasively and quantitatively.Part 3 Construction of deep learning models for rENE based on convolutional neural networksObjective: To construct a deep learning model of rENE based on convolutional neural networks.Methods: The included cases were the same as part 2.After the bp MRI sequences were jointly analyzed by two urologists,the cases were divided into rENE+ and rENE-groups.The images were segmented by 3D slicer software,while the program was later performed using Py Charm board with Python 3.8 software for image preprocessing and DL model building.To ensure a uniform resolution of all image data,the median size of spacing（0.4,0.4,4.5）was selected for resampling;and then all pixels were regularized using a z-score to ensure normal pixel distribution.The Py Torch 1.4 tool was introduced to build a layer-guided attention module,SCL,LL,and DL model construction.The accuracy of the model was evaluated using the F1 score;the ACC,SPR,SEN,and AUC were calculated to evaluate the model performance.Results: A deep learning model ENE-Net has been built,and the performance for predicting rENE was excellent.the F1 score,ACC,SPE,SEN,and AUC were 0.909±0.052,0.762±0.383,0.922±0.045 and 0.960±0.041 respectively.Conclusion: The DL model of rENE,ENE-Net,provides a good assessment of rENE and can be used to collect further cases in the future to complete iterations.