PSMA-targeted Precision Diagnosis Of Prostate Cancer

[Background]Prostate cancer is one of the malignant tumors that seriously threatens the life and health of men.In developed countries in the West,it has the highest incidence rate and the second highest mortality rate among males.In China,with the improvement of people’s living standards and the increase in average life expectancy,the incidence rate of prostate cancer has been increasing at a rate of 12.6%per year in the past decade,making it the fastest growing malignant tumor in China,seriously threatening the lives and health of the Chinese people.After radical prostatectomy in early-stage prostate cancer patients,the 5-year survival rate can reach more than 90%.However,in mid-to late-stage patients with distant metastasis,the 5-year survival rate is only about 33%.Therefore,accurate early diagnosis and timely treatment are crucial for improving the 5-year survival rate of prostate cancer patients.Prostate cancer lesions exhibit strong heterogeneity and can be classified based on the Gleason score(GS)into indolent prostate cancer(GS=3+3)and clinically significant prostate cancer(csPCa,GS>3+3).Conservative treatment measures such as active surveillance are recommended for the former,while radical treatments such as surgery or radiation therapy are adopted for the latter.Therefore,accurately distinguishing csPCa is an important clinical issue in the field of prostate cancer,which can prevent overdiagnosis and overtreatment of patients with indolent prostate cancer,and avoid underestimating the GS score of csPCa patients,delaying treatment.Since 2018,68Ga-PSMA PET/CT has gradually become popular worldwide for precise diagnosis of prostate cancer,and the 68Ga-PSMA PET guidelines proposed the need to identify the maximum standardized uptake value(SUVmax)diagnostic cutoff value for csPCa.In addition,the Louise Emmett team recently conducted a prospective multicenter PSMA PET/CT diagnostic efficacy study(PRIMARY study)for the diagnosis of csPC using 68Ga-PSMA PET/CT and proposed the PRIMARY-Score scoring method.This method combines anatomical localization(peripheral,central,or transition zone),prostate-specific membrane antigen(PSMA)activity pattern(none,diffuse,or focal),and high SUVmax(>12)to comprehensively assess the risk of csPCa.This method not only improves the diagnostic accuracy of PSMA PET/CT for csPCa,but also avoids errors caused by different PSMA ligands and PET equipment in different hospitals.However,Louise Emmett also pointed out that the effectiveness of this method still needs to be externally validated and confirmed.In the process of using and validating the PRIMARY-Score scoring method,we found that when SUVmax>12,the specificity of PRIMARY-Score for diagnosing csPCa is 100%.However,within the PRIMARY-Score range of 1-4,the false positive rates for scores 3 and 4 are as high as 62.1%and 23.9%,respectively,indicating that the PRIMARY-Score scoring method still needs to be further optimized.New technologies are contributing to further advance clinical problem solving and optimization.In recent years,diagnostic models based on artificial intelligence convolutional neural networks have achieved remarkable results in early detection of prostate cancer.Various convolutional neural networks have been used to assist in the pathological diagnosis of prostate cancer.By constructing artificial intelligence diagnostic models,quantitative evaluations of the benign and malignant nature of prostate cancer pathology and Gleason grading can be performed,providing methodological support for subsequent correlation analysis between PSMA-PET imaging and prostate cancer pathology.In clinical practice,prostate biopsy often underestimates csPC(GS>3+3)as clinically insignificant prostate cancer(GS=3+3),and cannot accurately predict csPCa based on clinical and pathological information alone.Previous studies have shown that the diagnostic efficiency can be improved by incorporating imaging or molecular information.Therefore,the correlation between PSMA expression and GS can be quantitatively analyzed using artificial intelligence convolutional neural networks,ultimately maximizing the diagnostic efficiency of csPCa based on PSMA PET/CT.In addition,Invasive cribriform carcinoma(ICC),Intraductal Carcinoma(IDC),and Ductal adenocarcinoma(DC)are special pathological subtypes of prostate cancer,which have similar molecular and prognostic characteristics.Currently,the standard diagnostic methods for prostate cancer,MRI and biopsy,have limitations in diagnosing ICC/IDC/DC,and the diagnostic efficacy and underlying mechanism of 68Ga-PSMA PET/CT for these lesions are not clear.This study focuses on the PSMA-PET molecular imaging technology for prostate cancer,combined with deep learning technology and single-cell sequencing technology,to address several clinical problems in the diagnosis of prostate cancer.[Objective]1)To determine the optimal SUVmax cutoff value for the diagnosis of csPC using 68Ga-PSMA PET/CT through a single-center retrospective cohort study and a prospective cohort validation.2)To externally validate the PRIMARY-Score diagnostic method proposed by an Australian team and analyze the diagnostic performance of 68Ga-PSMA PET/CT for prostate cancer through a single-center retrospective cohort study.3)To analyze the limitations of PRIMARY-Score through a multi-center retrospective cohort study and propose an optimized diagnostic method.To develop a 68GaPSMAPET/CT-based csPCa prediction model using contour plots to quantitatively explore the relationships among SUVmax,PRIMARY-Score,and csPCa risk,and fully utilize the advantages of PRIMARY-Score and SUVmax to achieve more accurate detection of csPCa.4)To construct a prostate cancer pathological diagnosis model and a Gleason pattern diagnosis model based on a deep learning convolutional neural network using our center’s data,providing methodological support for subsequent quantitative analysis of imaging and pathology.5)To explore the reasons for the limited diagnostic performance of SUVmax in the micrometer scale of lesions for patients with biopsy-proven GS=3+3 using the deep learning pathological diagnosis model developed and to propose a method to maximize the identification of patients with underestimated GS=3+3 through 68Ga-PSMA PET/CT and reduce unnecessary biopsies.6)To quantitatively explore the relationships among PSMA PET/CT SUVmax and several special pathological subtypes,including cribriform,ductal,and intraductal carcinoma,and investigate their potential molecular mechanisms using single-cell sequencing data of prostate cancer,through a single-center retrospective cohort study.[Methods]1)Patients who presented with symptoms and suspected prostate cancer(PCa)due to elevated prostate-specific antigen(PSA)levels and received treatment at the Department of Urology in Xijing Hospital between April 2017 and December 2019 were included in the analysis.All patients underwent 68Ga-PSMA PET/CT examination and prostate biopsy for pathological diagnosis.The primary observation endpoint of the study was csPC,and the secondary observation endpoint was the appearance of metastatic lesions.2)Patients who presented with symptoms and suspected PCa due to elevated PSA levels and received treatment at the Department of Urology in Xijing Hospital between June 2017 and April 2022 were included in the analysis.All patients underwent 68Ga-PSMAPET/CT examination and prostate biopsy for pathological diagnosis.The primary observation endpoint of the study was csPCa,and the primary evaluation indicator was the PRIMARY-Score.The diagnostic efficiency and clinical utility of the PRIMARY-Score for csPCa were evaluated by the receiver operating characteristic(ROC)curve and decision curve analysis(DCA),respectively.Cohen’s kappa coefficient was used to evaluate the inter-rater consistency of PRIMARY-Score.3)Patients who presented with symptoms and suspected PCa due to elevated PSA levels and received treatment at the Department of Urology in Xijing Hospital between June 2017 and June 2022 were divided into a training group and an internal validation group.Patients who presented with suspected PCa and received treatment at the Department of Urology in Nanjing Drum Tower Hospital affiliated with Nanjing Medical University or Tongji Hospital affiliated with Tongji Medical College of Huazhong University of Science and Technology between June 2021 and June 2022 were included in the external validation group.The primary observation endpoint of the study was csPCa.The primary evaluation indicators were the PRIMARY-Score and maximum standardized uptake value(SUVmax).Logistic regression analysis was used to determine the predictive factors,and restricted cubic spline curves were used to analyze the correlation between continuous variables and the risk of csPCa.Contour plots were used to illustrate the relationship between SUVmax and PRIMARY-Score and the risk of csPCa.The discrimination and calibration of the diagnostic method were evaluated using the area under the curve(AUC)and calibration curve,respectively,and the Akaike information criterion(AIC)was used to compare the model information loss.The diagnostic model was compared with SUVmax,PRIMARY-Score,ERSPC-RC3,and PBCG-RC models in the training set,internal validation set,and each subgroup,and internal validation was performed using 10-fold cross-validation(repeated 100 times).4)By performing HE staining and digital scanning on prostate tissue slices from patients in our center,we conducted performance testing on commonly used deep learning convolutional neural networks such as VGG16,ResNet50,MobileNet v2,Inception-v3,DenseNet,and Xception in terms of detection accuracy and inference speed,and built prostate cancer pathological diagnosis models and Gleason pattern diagnosis models.5)Prostate cancer patients with biopsy GS=3+3 who visited the Department of Urology at Xijing Hospital from June 2017 to June 2022 or the Drum Tower Hospital affiliated with Nanjing Medical University from January 2020 to January 2022 were included in the analysis.We used the previously constructed deep learning convolutional neural network pathological diagnosis model to analyze the reasons for the limited diagnostic performance of SUVmax in this population and developed a diagnostic method that maximizes the use of PSMA PET/CT to identify the underestimated population with biopsy GS=3+3 and reduce unnecessary biopsies.6)Patients who visited the Department of Urology at Xijing Hospital from June 2017 to June 2022 and were suspected of having prostate cancer based on symptoms and elevated PSA levels were included in the analysis.All patients underwent 68GaPSMA PET/CT examination,prostatectomy,and related pathological diagnosis.We analyzed the association model between SUVmax and ICC/IDC/DC lesion risks through restricted cubic spline curves and tested its nonlinearity.We tested saturation effects using smoothed curve plots and used the area under the curve(AUC)to evaluate the diagnostic performance of SUVmax for ICC/IDC/DC lesions.We explored the potential molecular mechanisms of PSMA PET/CT diagnosis of ICC/IDC/DC lesions through single-cell sequencing data.[Results]1)Through a retrospective development and prospective validation study at a single center,it was found that an optimal SUVmax cutoff value of 5.30 effectively distinguishes between csPCa and benign lesions,improving the accuracy of image interpretation.According to the ROC curve analysis,the optimal cut-off value for distinguishing csPC from benign lesions based on SUVmax was 5.30(sensitivity 85.85%,specificity 86.21%,AUC 0.893).This cutoff value achieved a sensitivity of 83.33%,specificity of 81.25%,positive predictive value of 92.11%,negative predictive value of 65.00%,and accuracy of 82.76%in the prospective validation cohort.In addition,it achieved a sensitivity of 76.19%,specificity of 81.25%,positive predictive value of 84.21%,negative predictive value of 72.22%,and accuracy of 78.38%in the same cohort.When combined with the presence of metastatic lesions,this cut-off value achieved an accuracy of 89.12%in all patients.Compared to manual image interpretation,the use of this cut-off value improved accuracy by 8.29%and reduced the number of ambiguous interpretations.SUVmax was strongly correlated with PSMA expression(r=0.831,P<0.001)and moderately correlated with GS(r=0.509,P<0.001).The PSMA expression levels and SUVmax values in csPCa patients were significantly higher than those in benign lesion patients(P<0.001).2)External validation studies have shown that the PRIMARY-Score has good diagnostic performance for csPCa and demonstrates minimal variability across different hospitals.However,further optimization is still necessary.Among the 346 patients included in the final analysis,170(49.1%)were confirmed to have csPCa based on biopsy pathology.The sensitivity,specificity,positive predictive value,and negative predictive value of PRIMARY-Score(1-2 vs.3-5)were 97.6%,38.1%,60.4%,and 94.4%,respectively.The AUC of PRIMARYScore(0.843;95%CI:0.803 to 0.883)did not differ significantly from that of SUVmax(0.846;95%CI:0.806 to 0.886;P=0.847),but was significantly higher than that of ERSPC-RC3(0.783;95%CI:0.735 to 0.831;P=0.042)and PBCG-RC(0.742;95%CI:0.690 to 0.794;P<0.001).Within the probability threshold range of 10%to 40%,PRIMARY-Score showed the highest net benefit,followed by SUVmax,ERSPC-RC3,and PBCG-RC.The inter-rater agreement for PRIMARY-Score 1-4 was 0.77(95%CI:0.70-0.84).3)Through a multicenter retrospective development and validation study,it was found that combining the PRIMARY-Score with SUVmax further improves the diagnostic performance of 68Ga-PSMA PET/CT for csPCa.The final form of the diagnostic model for csPC based on 68Ga-PSMA PET/CT is "Linear Predictor=10xPRIMARY-Score+3xSUVmax",with an AUC value of 0.8208(95%CI:0.7601-0.8815)in the training group and 0.7817(95%CI:0.6822-0.8812)in the internal validation group.The diagnostic accuracy of this model was superior to using only PRIMARY-Score and SUVmax in all groups.Based on the calculation of 10xPRIMARY-Score+3xSUVmax,patients can be classified into three risk groups with cutoff values of 40 and 60:≤40,>40 but ≤60,and>60,with csPCa positivity rates of 5.1%(95%CI:2.0%-12.5%),36.2%(95%CI:28.4%44.9%),and 74.2%(95%CI:62.6%-83.3%),respectively.4)A prostate cancer pathological model based on artificial intelligence and deep learning technology can be used for quantitative diagnostic analysis.MobileNet achieved the highest accuracy among six pre-trained convolutional neural networks.Moreover,MobileNet only took an average of 116 milliseconds to predict an input image,which was the shortest time among the six pre-trained convolutional neural networks.Using this diagnostic model,the ROC analysis of the tests performed on 61 cases at the Xijing Hospital showed an AUC value of 0.940.Under traditional methods,it takes about 5 minutes to evaluate a single case with biopsy pathology.With the help of this convolutional neural network model,pathologists can check the automatic diagnostic results in just 20 seconds.This approach realizes the quantitative evaluation of pathological features and rapid diagnosis of pathological results.5)A multicenter retrospective study showed that the diagnostic accuracy of prostate cancer biopsy in patients with a GS of 3+3 can be improved by using the diagnostic strategy of "10×PRIMARY-Score+2×SUVmax".For patients with GS=3+3 and SUVmax<12 undergoing transrectal ultrasound-guided biopsy,there is significant overlap in PSMA expression H-scores between benign glandular prostatic lesions/GP3/GP4 lesions,leading to overlapping SUVmax values between csPCa and non-csPCa lesions.The diagnostic model"10×PRIMARY-Score+2×SUVmax" has a significantly higher AUC(0.8359,95%CI:0.7233-0.9484)than PRIMARY-Score(0.7353,P=0.048),SUVmax(0.7256,P=0.009),the European Randomized Study of Screening for Prostate Cancer(ERSPC)risk calculator 3(0.6385,P=0.033),and the Prostate Biopsy Collaborative Group(PBCG)risk calculator(0.5468,P=0.001).Internal validation of this diagnostic model using 5-fold cross-validation with 1000 repetitions yielded an AUC of 0.8357(95%CI 0.8357-0.8358).The model demonstrated superior diagnostic performance across different subgroups.6)SUVmax is an independent risk factor for invasive cribriform/ductal carcinoma and ductal adenocarcinoma.The upregulation of PSMA expression due to PTEN loss activating CBP/p300/HOXB13 is a potential molecular mechanism for this phenomenon.ICC/IDC/DC is closely associated with high GS scores(GG1-2 vs.GG3-5,P<0.001)and lymph node metastasis(pN0 vs.pN1,P=0.004).Moreover,the SUVmax values of ICC/IDC/DC-positive lesions are significantly higher(based on a patient-level analysis,14.69 vs.5.88,P<0.001).There is a nonlinear relationship(nonlinear P<0.001)between SUVmax and the risk of having ICC/IDC/DC features,with a saturation point of 13.70(based on a patient-level analysis).Before the saturation point,each 1-unit increase in SUVmax is associated with a 41.6%increase in the risk of having ICC/IDC/DC features(OR=1.416,95%CI:1.231-1.630;P<0.001).After the saturation point,there is no significant correlation between increasing SUVmax and the risk of having ICC/IDC/DC features.The ROC curve analysis shows that the AUC value of SUVmax for ICC/IDC/DC lesions is 0.835(based on a patient-level analysis,95%CI:0.755-0.897,P<0.001).Protein-level immunohistochemical analysis and single-cell transcriptomic analysis suggest that PTEN loss may lead to PSMA overexpression and high SUVmax,and CBP/p300/HOXB13 upregulation is a potential molecular mechanism.[Conclusions]1)This study established and prospectively validated an optimal SUVmax cutoff value of 5.30 for accurately identifying csPCa.2)The PRIMARY-Score scoring system showed good diagnostic performance and clinical utility in external validation,reducing unnecessary biopsies and serving as an effective supplement to existing csPCa risk prediction models.3)The diagnostic model combining PRIMARY-Score and SUVmax is a further optimization of the PRIMARY-Score scoring system,maximizing the use of information from 68Ga-PSMA PET/CT for assessing csPCa risk.4)The prostate cancer diagnostic model based on deep learning convolutional neural networks can be used for quantitative evaluation of pathological features,assisting in the correlation analysis of imaging and pathology.5)For patients with biopsy GS=3+3 and SUVmax<12,the "10×PRIMARY-Score+2×SUVmax" method efficiently predicts csPCa,especially for patients with"anterior stromal escape lesions".6)68Ga-PSMA PET/CT SUVmax is helpful in identifying ICC/IDC/DC lesions,and PTEN loss in lesions is a potential molecular mechanism for upregulation of PSMA expression and SUVmax through CBP/p300/HOXB13.