Deep Neural Network Analysis Employing Diffusion Basis Spectrum Imaging Metrics As Classifiers For Differential Diagnosis Of Prostate Cancer And Prediction Of Pathological Grades

Part I.The application of DBSI-based deep neural network(DHI)in the differential diagnosis of prostate cancerPurposesMultiparametric magnetic resonance imaging(MP-MRI)is currently considered as one of the most excellent imaging methods for the diagnosis of prostate cancer(PCa).In recent years,MP-MRI has played an important role in the early diagnosis and targeted biopsy of prostate cancer.With the widely application of MP-MRI,more and more early prostate cancers have been diagnosed.The combination of targeted biopsy also further improves the accuracy of our judgment on the index lesion of prostate cancer,allowing us to detect more clinically significant prostate cancers,while avoiding the detection of clinically insignificant cancers.But through systematic biopsy,we know that even in patients with negative MRI,there are still about 15% of patients with clinically significant prostate cancer,and about 12% of patients with PI-RADS scores of 5 are non-prostate cancer.Therefore,we urgently need a non-invasive imaging method that can more accurately differentiate and diagnose prostate cancer.In MP-MRI,diffusion weighted imaging(DWI)is the most important sequences for evaluating the differential diagnosis of prostate cancer.Yet the DWI sequence we routinely use is designed according to a single index model,which means it models the tissue as a whole,and does not distinguish whether the restriction of the diffusion of water molecules comes from the isotropic part or the anisotropic part.And it does not further distinguish the isotropic diffusion part,which leads to the lower specificity and sensitivity of the diagnosis of DWI.Therefore,we proposed a new diffusion-weighted imaging model-diffusion based spectrum imaging(DBSI),which innovatively sepearates the diffusion of water molecules in tissues into anisotropic and isotropic parts.The anisotropic part is expressed as the fiber fraction,and the quantitative parameter index of this part is used to model the fiber composition in the prostate.The isotropic diffusion part is further divided into the highly restricted fraction,restricted diffusion fraction,hindered diffusion fraction and free diffusion fraction.according to the extent of the restricted diffusion of water molecules.Highly restricted frraction models the highly restricted diffusion part of water molecules in the inflammatory cells.The restricted fraction models the restricted part of the diffusion of water molecules in tumor cells.The hindered fraction models the water molecules in the interstitial cells.and the free fraction models the diffusion of water molecules in the normal ductal acini.This study is to use DBSI combined with deep neural network learning classifier(named diffusion histology imaging,DHI)to evaluate its efficacy in differential diagnosis of prostate cancer using large pathological slices after radical prostatectomy as the gold standard.Materials and methodsProspective and continuous recruitment of patients with clinically suspected prostate cancer from March 2015 to August 2017.Inclusion criteria:(1)PSA elevation is greater than 4ng/ml;(2)abnormal echo on ultrasound examination;(3)positive findings on digital rectal examination.Exclusion criteria:(1)Patients who have received radiotherapy or endocrine therapy in the past;(2)Those who have contraindications to magnetic resonance examination;(3)Those who suffer from claustrophobia.All enrolled patients received multi-parameter MR and DBSI scans at the same time.Multi-parameter MR was evaluated according to the PI-RADS 2 scoring standard.Patients with PI-RADS ? 3 were selected for system + targeted puncture.After biopsy,suitable patients were selected for radical prostatectomy.For patients undergoing radical surgery,follow the postoperative Large pathological sections delineate cancer foci,select patients with PI-RADS ? 3 and receive targeted biopsy but the result is negative as the benign focus control group,and delineate the benign area according to the results of targeted biopsy.A supervised deep neural network(DNN)is used to construct the classifier.A total of 22 DBSI diffusion metrics were used to construct the prediction model.Supervised machine learning based on deep neural networks was used to conduct receiver operating curve(ROC)analysis to evaluate the ability of DBSI to distinguish PCa from other pathologies or tissue types.We calculated area under the curve(AUC),sensitivity and specificity values.All statistical analyses were perrformed using Python(version 3.8)packages Tensor Flow(version 2.0),Scikit-learn and Scipy.ResultsA total of 243 people(mean age 65.8 years old,mean PSA 25.4ng/ml)were enrolled in the study,96 patients(average age 69.3 years old,average PSA 27.9ng/ml)were finally diagnosed with prostate cancer through pathology,and the biopsy was positive.The positive rate was 25.8%(391/1515).Ninety-two patients underwent radical prostatectomy,and the specimens of the patients who underwent radical resection were handed over to experienced urological pathologists for pathological diagnosis,and pathological risk classification machine for pathological clinical staging.And outline the cancer foci according to the pathological large section.A total of 54 patients(mean age 65 years,mean PSA11.5ng/ml)who underwent biopsy were diagnosed with benign prostatic hyperplasia or prostatic inflammation.We selected 54 patients as the benign control group based on the results of the targeted biopsy.Among them,93 patients(mean age 62 years,average PSA9.8ng/ml)did not receive biopsy due to PI-RADS score <3 and PSA less than 10 ng/ml,so dynamic follow-up was used.Through deep neural network learning,the diagnostic efficiency of DBSI in identifying prostate cancer in patients at the voxel level is as follows:the AUC,sensitivity and specificity to distinguish PCa from benign peripheral zone tissue is 0.995,94.8%,and 98.3%,respectively;The AUC,sensitivity and specificity of DHI to distinguish PCa from the benign transition zone was 0.985,93.9%,and 94.7%,respectively.The AUC,sensitivity and specificity of DHI to distinguish PCa from benign prostate tissue were 0.998,97.7%,and 97.8%,respectively.ConclusionDBSI-based deep neural network(DHI)has high diagnostic efficiency in the identification of benign and malignant prostate diseases.It can accurately differentiate and diagnose prostate cancer,benign peripheral prostate disease,benign transitional zone disease and benign prostate disease,and is targeted for follow-up Puncture and precision treatment are meaningful.Part II.The application of DBSI-based deep neural network in predicting the pathological grading of prostate cancerPurposesThe most important part of the diagnosis of prostate cancer(PCa)is to assess its pathological risk grades.Because the prognosis of PCa with different tumor grades is significantly different,most of the clinically non-significant cancers only require dynamic follow-up or active monitoring.And clinically significant cancer usually requires more aggressive treatment.The gold standard for the diagnosis of PCa is still transrectal or transperineal systematic biopsy.However,the diagnosis of this method often differs from the postoperative pathological score.Therefore,it is particularly important to accurately predict the pathological grading of PCa patients before surgery.Previous studies have shown that DWI has a certain application value in predicting the pathological grading of patients.Our research results suggested that diffusion basis spectrum imaging(DBSI)is more advantageous than conventional DWI in the differential diagnosis of PCa.Therefore,we speculate that DBSI has more advantages than conventional DWI in predicting the pathological grade of patients.This study prepares to use the deep neural network method combined with DBSI(named diffusion histology imaging,DHI)to evaluate its application value in predicting the pathological grading of patients.Materials and MethodsProspective and continuous recruitment of patients with clinically suspected PCa from March 2015 to August 2017.Inclusion criteria:(1)PSA elevation is greater than 4 ng/ml;(2)abnormal echo on ultrasound examination;(3)positive findings on digital rectal examination.Exclusion criteria:(1)Patients who have received radiotherapy or endocrine therapy in the past;(2)Those who have contraindications to MRI examination;(3)Those who suffer from claustrophobia.All enrolled patients received multi-parameteric MRI(MP-MRI)and DBSI scans at the same time.MP-MRI was evaluated according to the PI-RADS 2 scoring standard.Patients with PI-RADS ? 3 were selected for systemic +targeted puncture.After biopsy,suitable patients were selected for radical prostatectomy,and patients who had undergone radical surgery were selected for postoperative diagnosis.The pathological section was outlined by a urological pathologist and scored according to the International Society of Urological Pathology(ISUP)grade group 1-5.A supervised deep neural network(DNN)is used to construct the classifier.A total of 22 DBSI diffusion metrics were used to construct the prediction model.Supervised learning based on DNN was used to calculate receiver operating curve(ROC)analysis to evaluate DBSI's predictive ability for differentiating ISUP pathological grades.To evaluate the accuracy of its prediction,we calculated the area under the curve(AUC),sensitivity and specificity values.All calculations were obtained using Python(version 3.8)packages Tensor Flow(version 2.0),Scikit-learn and Scipy.ResultsA total of 243 people(mean age 65.8 years,mean PSA 25.4 ng/ml)were enrolled in the study,of which 96 patients(average age 69.3 years,average PSA 27.9ng/ml)were finally diagnosed with PCa through pathology,and the biopsy was positive.The positive rate was 25.8%(391/1515).Ninety-two patients underwent radical resection of prostate cancer.The specimens of patients who underwent radical resection were handed over to experienced urological pathologists for pathological diagnosis based on whole mount section.The PCa grading was evaluated according to ISUP PCa pathological grading of Gleason grade groups(GGG).Through deep neural network learning,the results of DHI in predicting the pathological classification of PCa ISUP were as follows: the overall accuracy of ISUP classification based on voxel level was 91.4%.The classification recall rates of GGG 1,GGG 2,GGG 3,GGG 4,and GGG 5 are 84.0%,89.6%,91.5%,89.8%,and 97.8%,respectively.AUC values reached 0.989,0.968,0.967,0.972,0.978,respectively.The sensitivity and specificity values of diagnosis were 96.7% and 93.6%,94.0% and 91.5%,90.7% and 91.5%,93.7% and 90.4%,92.7% and 92.3%,respectively.ConclusionDBSI-based deep neural network(DHI)can accurately predict the pathological grading of PCa patients and provide strong support for subsequent precision treatment.It is a non-invasive imaging method for assessing the pathological grading of PCa and has good clinical results.Part III DBSI-based deep neural network in the differential diagnosis of prostate cancer and prediction of pathological grading in ex vivo prostate specimensPurposesNon-invasive differential diagnosis and predictive pathological grading of prostate have always been the focus of imaging diagnosis.Currently,multi-parametric magnetic resonance imaging(MP-MRI)is recognized as the best noninvasive differential diagnosis and predictive imaging method for prostate cancer(PCa)pathological grading.But benign diseases such as prostatic inflammation and prostatic stromal hyperplasia can mimic the imaging manifestations of prostate cancer under MP-MRI.In terms of predicting pathological grades,although MP-MRI has certain application value,the degree of intersection between different pathological grades is obvious,and it is difficult to make accurate predictions.The main reason is that the conventional DWI sequence does not distinguish the prostate tissue well.Diffusion basis spectrum imaging(DBSI),a new diffusion-weighted imaging model proposed by us,innovatively distinguishes the diffusion-limited part of water molecules in the tissue into anisotropic and isotropic portions.The isotropic portions are further differentiated according to the degree of restriction of water molecules in the prostate tissue.In the previous PCa patient study,we have concluded that conventional MP-MRI is the most important aim in the differential diagnosis of prostate cancer or in predicting the pathological grade of the moderate diagnosis is to assess its pathological risk grade.The prognosis of prostate cancer of different risk grades is significantly different.Most of the clinically non-significant cancers only require dynamic follow-up or active monitoring,which will not endanger the life of the patient,while clinically significant cancers require more active treatment.We focus on prostate cancer.The gold standard for diagnosis is still system biopsy through the rectum or perineum,but the pathological scores after diagnosis by this method are often different from those after surgery.Therefore,it is particularly important to accurately predict the pathological grading of prostate cancer patients before surgery.The previous related research It has been shown that DWI has certain application value in predicting the pathological grading of patients.Our research results also confirm that DBSI is more advantageous than conventional DWI in the differential diagnosis of prostate cancer and predicting pathological grade.Therefore,we will further evaluate the application value of DBSI in the differential diagnosis of prostate cancer and the prediction of pathological grade in ex vivo prostate tissue samples.In this study,deep neural network learning was used as a classifier combined with DBSI to evaluate its application value in differential diagnosis and prediction of pathological grading of prostate cancer in ex vivo prostate tissue samples.Materials and MethodsFrom March 2015 to August 2017,97 prostate cancer specimens(including 9 from Shanghai Changhai Hospital and 10 from Washington University)after radical prostatectomy were selected for in vitro standard ultra-high field MRI scans.Changhai Hospital used the 9.7T MRI scanner to scan the tissue,and the Washington University used a 4.7T MRI scanner.Imaging sequence includes conventional T2 WI,DWI and DBSI.After MRI scan,the specimens were sectioned and stained with H&E.The prostate cancer area,prostatic hyperplasia tissue,and prostatic inflammation tissue were delineated by urological pathologists,and the prostate cancer was graded.We used Python-based Tensor Flow version 2.0 software to build a deep neural network to evaluate the ability of DBSI in the differential diagnosis of prostate cancer and predict the pathological grade of ISUP.We used receiver operating curve(ROC)analysis to evaluate the differential diagnosis ability of DBSI for different prostate diseases and the predictive ability of ISUP pathological grading.We calcualted the prediction accuracy,the area under the curve(AUC),sensitivity and specificity values.All the statiscial metrics and curves were calculated based on the Python packages of Sciki-learn and Scipy.Results97 prostate specimens were scanned by MRI and were sectioned into 97 HE-stained pathological sections by pathology slicer.Total 48 prostate cancer lesions,39 prostate hyperplasia nodules and 40 stromal prostate hyperplasia areas and 76 benign peripheral zone tissue were drawn on the HE sections by an experienced pathologist(Professor Yu Yongwei).For prostatectomy specimens,DHI distinguishes prostate cancer from benign peripheral areas with AUC of 0.949,sensitivity of 87.5%,and specificity of 88.4%;DHI distinguishes prostate cancer from stroma areas with AUC of 0.928,sensitivity of 86.6%,and specificity of 84.9%;DHI distinguishes prostate cancer from benign prostatic hyperplasia with AUC of 0.900,sensitivity of 82.6%,and specificity of 81.6%;DHI distinguishes prostate cancer from benign prostatic hyperplasia with AUC of 0.911,sensitivity of 84.2%,and specificity of 83.0%.The overall accuracy of DHI for ISUP Gleason grade grroups of prostatectomy specimens was 72.1%.The classification accuracy based on voxel level for GGG 1,2,3,and 5 are 70.1%,76.0%,75.6%,and 71.1%,respectively.ConclusionDBSI-based deep neural network(DHI)can accurately differentiate and diagnose prostate cancer and benign peripheral zone of prostate and stromal hyperplasia prostate tissue in ex vivo prostate specimens.It can also predict the ISUP pathological grading of prostate cancer tissue with moderate accuracy.