Value Of MRI And Its Intelligent Post-processing In Prostate Cancer With PI-RADS 3

Objective: To investigate the feasibility and efficacy of MRI and its intelligent post-processing in the detection of prostate cancer(PCa)and clinically significant prostate cancer(cs PCa)in PI-RADS 3 lesions.Methods: The hospital’s ethics committee approved the study,exempting patients from providing informed consent.A retrospective analysis was performed on patients with prostate imaging-reporting and data system(PI-RADS)category 3 lesions who received multi-parametric MRI or bi-parametric MRI between January 2015 and December 2021.Biopsy,transurethral resection of the prostate,or radical prostatectomy specimens were used to make the pathological diagnosis,and the lesions were divided into three groups: benign(group 1),non-cs PCa(Gleason score ≤ 3+3)(group 2),and cs PCa(Gleason score ≥ 3+4)(group 3).(1)Independent risk factors associated with PCa and cs PCa in PI-RADS 3 lesions were analyzed among indicators of age,prostate-specific antigen,prostate-specific antigen density,apparent diffusion coefficient(ADC),T2 WI signal intensity(T2WISI),ADC heterogeneity index,T2 WI signal intensity heterogeneity index(T2WISIcv),ADC density,prostate-specific antigen lesion volume density,ADC lesion volume density,and the diagnostic value of independent risk factors was evaluated by receiver operator characteristic(ROC)curve.(2)T2WI and ADC images were selected to delineate intra-lesional area of interest of PI-RADS 3 lesions,which were then automatically expanded to 3mm around the lesion to obtain peri-lesional area of interest,and the radiomics features were extracted.The dataset was divided into a training set and a test set in a 6:4 ratio.The feature dimension was reduced by T-test,correlation de-redundancy and least absolute shrinkage and selection operator regression.Logistic,support vector machine(SVM),K-nearest neighbor(KNN),naive Bayes(NB),random forest(RF)and gradient boosting machine(GBM)models were constructed and validated in the test set.(3)T2WI images of PI-RADS 3 lesions were saved in JPEG format and divided into a training set and a test set in an 8:2 ratio.The deep learning framework based on Res Net-18 was used for model training and the prediction efficiency of the model was validated in the test set;The semantic features of lesions and the image region targeted by the model are visualized.Results: A total of 202 patients were included in the study,with 133 patients in group 1,25 patients in group 2 and 44 patients in group 3.There were no significant differences in baseline characteristics among the three groups(P > 0.05),including age,prostate-specific antigen,and lesion location.(1)The correlation of volume of PI-RADS 3 lesions measured by the formula method and segmentation method was 0.9,thus the formula method was chosen to calculate the prostate volume and PI-RADS 3 lesion volume.Univariate and multivariate analysis showed that T2 WISI and ADC heterogeneity index were independent risk factors for the diagnosis of PCa in PI-RADS 3 lesions.ROC curve analysis showed that a combination of T2 WISI and ADC heterogeneity index to diagnose PCa in PI-RADS 3 lesions yields an area under the curve(AUC)of 0.68.Univariate and multivariate analysis showed that ADC was an independent risk factor for the diagnosis of cs PCa in PI-RADS 3 lesions.ROC curve analysis showed that using ADC to diagnose cs PCa in PI-RADS 3 lesions yields an AUC of 0.65.(2)1037 radiomics features were extracted from the intra-lesional and peri-lesional regions of interest.In the diagnosis of PCa,33 features were selected.The naive Bayes model had the highest diagnostic efficacy,with an AUC of 0.765 in the test set.In the analysis of cs PCa,27 features were selected.The K nearest neighbor model had the highest diagnostic efficacy,with an AUC of 0.65 in the test set.Decision curve analysis showed that Radscore could benefit patients.(3)A total of 423 images of benign prostate lesions,158 images of non-cs PCa and 273 images of cs PCa were included,with 685 images used for training the model and 169 images used for testing.The best accuracy of the model trained by Res Net-18 in the testing set was 0.787.The average precision of the deep learning model in predicting benign prostate disease,non-cs PCa and cs PCa was 0.882,0.681 and 0.851,and the AUC was 0.875,0.89 and 0.929,respectively.Semantic feature analysis revealed that cs PCa and benign prostate lesions were highly distinguishable.The class activation map demonstrates that the deep learning model can focus on the prostate region or the location of the lesion.Conclusions: The T2 WISI and ADC heterogeneity index can be used to diagnose PCa in PI-RADS 3 lesions,and the ADC can be used to diagnose cs PCa in PI-RADS 3 lesions.Based on the intra-lesional and peri-lesional radiomics features of T2 WI and ADC images, the machine learning model could be used to diagnose PCa and cs PCa in PI-RADS 3 lesions.The best models are NB and KNN models,respectively.The deep learning model based on the Res Net-18 architecture can achieve the intelligent classification of PI-RADS 3 lesions.Part Ⅰ: Diagnosis of PCa and cs PCa in PI-RADS 3 lesions based on clinical-imaging indicatorsObjective: To explore the feasibility and efficacy of clinical imaging indicators in the diagnosis of PCa and cs PCa in PI-RADS 3 lesions.Methods: The study was approved by the hospital ethics committee and the informed consent of patients was exempted.A retrospective analysis was performed on the lesions that were diagnosed as PI-RADS 3 by multi-parametric MRI or bi-parametric MRI from January 2015 to December 2021.Pathological diagnosis of biopsy and transurethral resection of prostate or radical prostatectomy specimens was taken as the gold standard,and the lesions were divided into benign,non-cs PCa and cs PCa groups.ADC and T2 WISI was measured on the slice showing the maximum diameter of the lesion at the post-processing workstation,and the heterogeneity index was calculated.The volume of the lesion was measured by the formula and segmentation methods,and the correlation between them was calculated.The prostate-specific antigen density,ADC density,prostate-specific antigen lesion volume density and ADC value lesion volume density were calculated.Univariate and multivariate analyses were used to screen risk factors associated with PCa(group 2 + group 3 vs group 1)and cs PCa(group 3 vs group 1 + group 2).The efficacy of independent risk factors in the diagnosis of PCa and cs PCa in PI-RADS 3 lesions was evaluated using the ROC curve.Results: A total of 202 patients were included,including 133 patients with benign prostate disease,25 patients with non-cs PCa and 44 patients with cs PCa.There were no statistically significant differences among the three groups in baseline characteristics including age,prostate-specific antigen and lesion location(P>0.05).The correlation of the volume of PI-RADS 3 lesions measured by the formula method and segmentation method was 0.9,thus the formula method was chosen to calculate the prostate volume and PI-RADS 3 lesion volume.Univariate and multivariate analysis showed that T2 WISI and ADC heterogeneity index were independent risk factors for the PI-RADS 3 lesions to be diagnosed as PCa.ROC analysis showed that the AUC was 0.68 with the combination of T2 WISI and ADC heterogeneity index,and the analysis of the decision curve showed that the patients could benefit;Univariate and multivariate analysis showed that ADC was an independent risk factor for PI-RADS 3-point lesions to be diagnosed as cs PCa.ROC analysis showed that the AUC for ADC to diagnose cs PCa was 0.65,and the decision curve analysis showed that the patients could benefit.Conclusions: T2 WISI and coefficient of variation of ADC can be used to diagnose PCa in PI-RADS 3 lesions,and ADC can be used to diagnose cs PCa in PI-RADS 3 lesions.Part Ⅱ: Machine learning models based on intra-lesional and peri-lesional radiomics features to predict PCa and cs PCa in PI-RADS 3 lesionsObjective: To explore the feasibility and efficiency of the machine learning model based on the intra-lesional and peri-lesional radiomics features in the prediction of PCa and cs PCa in PI-RADS 3 lesions.Methods: Retrospective analysis was performed on lesions diagnosed with PI-RADS 3 by multi-parametric MRI or bi-parametric MRI from January 2015 to December 2021.Pathological diagnosis on biopsy and transurethral resection of prostate or radical prostatectomy specimens were used as the gold standard,and the lesions were divided into a benign group,non-cs PCa group and cs PCa group.In 3D-Slicer software,the intra-lesional area of interest of the PI-RADS 3 lesions in T2 WI and ADC images were delineated,and the peri-lesional area of interest of the PI-RADS 3 lesions were achieved automatically extending to 3mm outside the lesions.The radiomics features of the lesions were extracted by the Pyradiomics package.Feature screening was performed by T-test,and the training set and validation set were divided according to 6:4.Then the feature dimension was reduced by correlation de-redundancy and least absolute shrinkage and selection operator(LASSO)regression.Logistic,SVM,KNN,NB,RF and GBM models were constructed and used for the prediction of PCa(group 2 + group 3 vs group 1)and cs PCa(group 3 vs group 1 + group 2)in the test set.ROC analysis was used to evaluate the diagnostic effectiveness of the models.Radscore was calculated for each patient and clinical benefit was assessed using decision curves.Results: A total of 202 patients were included,including 133 patients in group 1,25 patients in group 2,and 44 patients in group 3.1037 features were extracted from the intra-lesional and peri-lesional areas of interest.In the diagnosis of PCa,1483 features were selected by T-test,371 features were left after correlation de-redundancy analysis,and 33 features were selected by LASSO regression.It was found that the diagnostic efficiency of the NB model was the highest,with an AUC reaching 0.765.In the diagnosis of cs PCa,1076 features were selected by T-test,267 features were left after correlation de-redundancy analysis and 27 features were selected by LASSO regression.KNN model had the highest diagnostic efficiency,with an AUC of 0.65.The decision curve analysis of radiomics radscore can benefit patients.Conclusions: Machine learning models based on intra-lesional and peri-lesional radiomics of PI-RADS 3 lesions could be used to predict PCa and cs PCa.The best models are NB and KNN,respectively,with areas under the curve of 0.765 and 0.65.Part ⅡI: Intelligent classification of PI-RADS 3 lesions by deep learningObjective: To explore the intelligent classification and prediction efficacy of the deep learning model for benign prostate lesions,non-cs PCa and cs PCa in PI-RADS 3 lesions.Methods: From January 2015 to December 2021,lesions diagnosed with PI-RADS 3 by multi-parametric MRI or bi-parametric MRI were retrospectively included.They were classified as benign,non-cs PCa,and cs PCa according to the pathologic results of the biopsy,transurethral resection or radical prostatectomy specimens.T2 WI images of the lesions were saved in JPEG format and were divided into a training set and a test set according to 8:2;The preprocessing of the training set included scaling,clipping,image enhancement,tensor conversion and normalization.The preprocessing of the test set included scaling,clipping,tensor conversion and normalization.Res Net-18 architecture was used for model training,and all layers were trained.The GPU used was Tesla T4 and the computing device architecture version was v11.2.All statistical analyses were performed using Python open-source libraries such as numpy,pandas,matplotlib,seaborn,plotly requests,tqdm,opencv-python,pillow,wandb,etc.The trained model was saved as a pth file and evaluated in the test set.The ROC curve was used to evaluate the predictive effectiveness of the model.T-NSE was used for image semantic feature visualization.The class activation mapping was used to visualize the area focused by the model.Results: A total of 423 benign prostate lesion images,158 non-cs PCa images and 273 cs PCa images were included.A dataset of 685 images was used for training the deep learning model,while 169 images were used for testing.The model achieved an accuracy of 0.787 on the test set.The average precision in predicting benign prostate disease,non-cs PCa and cs PCa were 0.882,0.681 and 0.851,and the AUC were 0.875,0.89 and 0.929,respectively,for the three categories.Semantic feature analysis demonstrated a strong classification separability between cs PCa and benign diseases.Additionally,the class activation map showed that the deep learning model can focus on the area of the prostate or the location of PI-RADS 3 lesions.Conclusions: Deep learning model based on Res Net-18 architecture can realize intelligent classification of PI-RADS 3 lesions.