Prostate cancer (PCa) is the second most common malignant tumor in men worldwide and the fifth leading cause of death globally (1, 2). In recent years, with changes in lifestyle and the acceleration of aging, the incidence of PCa in the world also has increased year by year (1, 3). Zattoni classified prostate cancer into clinically significant prostate cancer (CS-PCa) according to its Gleason score (GS) and tumor invasion range, and they compared it to non-clinically significant PCa (Non-CS-PCa) (4). Non-CS-PCa has low aggressiveness and slow progression, and more follow-up and active monitoring are used without intervention treatment (5). CS-PCa is highly invasive and progresses rapidly. If not actively treated, it can lead to systemic metastasis, causing serious harm to patients’ lives and quality of life (6). Data show that at the initial diagnosis, more than half of PCa patients already have bone metastases, accompanied by bone pain, pathological fractures, movement disorders, spinal cord compression, and other complications (7). Hormonal therapy, radiotherapy, chemotherapy, and surgical treatment should be chosen according to the stage of the illness (8). The incidence of CS-PCa is higher in older men and men with no family history (9). Early detection of CS-PCa is of profound significance for the effective treatment of patients.

The difficulty with diagnosing and treating prostate cancer lies in the different biological behaviors of the tumors, and the characteristics of prostate cancer are multifocal, microfocal, and heterogeneous (10). Currently, a systematic ultrasound-guided biopsy is the main method for diagnosing prostate cancer. However, it can miss CS-PCa and overdiagnose Non-CS-PCa (11–13).

The clinical value of prostate-specific antigen (PSA) as a marker of prostate tumor has been recognized, but the high false positive rate for PSA can make it difficult to accurately diagnose PCa (14). Currently, researchers are exploring non-invasive imaging techniques to increase the accuracy of PCa diagnosis. In auxiliary imaging examinations, multi-parameter NMR has shown obvious advantages (15). Multimodal ultrasonography (MUS) refers to the combined application of transmittal color ultrasound (TRUS), transrectal real-time elastography (TRTE), and transrectal contrast enhanced ultrasound (TR-CEUS) technology to perform diagnostic rectal contrast enhanced ultrasound on diseases. Compared to using any of these technologies alone, a combined examination has been shown to improve the diagnosis rate (16). As the availability of transmittal MUS and multiparameter magnetic resonance in prostate cancer detection continues to improve, more studies are reporting the practicability of combining serum biomarkers with imaging evaluation. In order to improve the specificity of CS-PCa detection and avoid over-diagnosis of non-CS-PCA, a CS-PCa machine learning model was established and predicted by combining MUS and PSA-related risk indicators.

Depending on the applied prostate-specific antigen (PSA) cutoff as a trigger for a prostate biopsy, PSA can be a highly sensitive marker for prostate cancer (PCa) screening. However, up to 20% of clinically significant PCa (CS-PCa) can be missed on the first systematic TRUS-guided prostate biopsy in men with low PSA values (30–32). The relevant indicators of PSA should exhibit a certain value for predicting CS-PCa. In this study, unary logistic regression analysis was performed to determine the positive correlation of tPSA, fPSA, f/tPSA, PSAD, and other factors with CS-PCa (P< 0.001), while the fPSA/tPSA ratio was negatively correlated with CS-PCa (P< 0.001). The results obtained agree with those of Zhou (33). PSAD is an indicator that reflects the PSA per unit volume of the prostate, so it shows higher sensitivity and specificity than PSA (34). Studies have highlighted that PSAD is an important factor for predicting the postoperative Gleason score (GS), especially when GS< 7. The combined age of PSAD is of great value for the increase in GS after an operation (35). Hansen found that a PSAD cutoff value > 0.15 ng/mL/mL significantly improved the positive predictive value (PPV) (33%) for csPCa when performing targeted biopsies for PI-RADS 3 lesions (36). Venderink reported that using a PSAD cutoff value ≥ 0.15 ng/mL/mL for patients with PI-RADS 3 lesions resulted in 42% of the patients avoiding targeted biopsies, and 6% of cases of csPCa being missed (37). In this study, the age of the CS-PCa group was higher than that of the Non-CS-PCa group (75 vs. 73 years, P = 0.017), and the PSAD of the CS-PCa group was significantly higher than that of the Non-CS-PCa group (0.55 vs. 0.17, P< 0.001). Some studies have shown that prostate health index (PHI) is the best PSA-derived biomarker to predict CS-PCa in men with PI-RADS 3 lesions in a cognitive MRI-TRUS fusion targeted biopsy. PHI exhibits the highest prediction performance for CS-PCa with an AUC of 0.884, which is higher than that of PSAD. Targeted biopsies show more CS-PCa and less Non-CS-PCa than systematic biopsies (38).

Prostate MRI has developed into an important tool for the management of PCa. Prostate MRI is recommended as the first-line screening method for patients with a clinical suspicion of prostate cancer (39). The Prostate Imaging-Reporting and Data System (PI-RADS) represents a comprehensive set of guidelines, standardized observations, and lexicon, which aim to stratify the probability of clinically significant prostate cancer (csPCa) for MRI (40). In this study, the detection rates of CS-PCa and Non-CS-PCa were statistically analyzed by multimodal ultrasound scoring systems such as the 2D score, electrograph score, and CEUS score. The results showed that the differences were statistically significant (P< 0.001). The tumors of the Non-CS-PCa group showed high differentiation, low invasiveness, and relatively unobvious malignant signs in two-dimensional ultrasound images, whereas the tumors of the CS-PCa group showed low differentiation, high invasiveness, and proliferation of a large number of tumor cells, which may show an obvious space-occupying effect to form nodules, and may also involve the prostatic capsule and adjacent organs. PCa with different GS also exhibited different sonogram features. This resulted in a difference in the multimodal ultrasound scores of the Non-CS-PCa and CS-PCa groups. At present, the most widely used method for the clinical diagnosis of PCa is an ultrasound-guided transmittal prostate biopsy. However, the main limitations of this method are a large number of false negatives, the frequent failure to detect CS-PCa, and excessive puncture, which can cause misdiagnosis, missed diagnosis, and an increase in the incidence of complications after the examination. Grey used mpMRI to predict prostate biopsy results, and csPCa (ROC) curve analysis showed that the area under the curve (AUC) was 0.89, while the negative predictive value (NPV) of the PI-RADS Score 2 was 0.9814. However, the positive predictive value (PPV) of csPCa was 0.49, and a small number of patients with PI-RADS scores of 2 had CSPCa (41). The PROMIS trial was designed to evaluate the utility of mpMRI as a triage test in biopsy-naive patients to avoid unnecessary TRUS biopsy. The diagnostic accuracy of mpMRI and TRUS-guided biopsy was compared using transparency template prostate mapping as the reference standard. Ahmed reported that using mpMRI as a triage test resulted in avoiding 27% of primary biopsies, reducing the detection rate of clinically insignificant cancer by 5%, and improving the detection rate of csPCa by 18% (42). Wagner reported that the AUC of the model for PCa and csPCa prediction in patients with negative prostate MRI was 0.80 and 0.87, respectively, which showed that the number of biopsies in patients with negative MRI examination could be significantly reduced without a significant loss of CS-PCa detection (43).

Endemics can quantitatively analyze imaging data to noninvasively evaluate the biological behavior of tumors and has been widely used in the diagnosis, invasiveness assessment, and clinical decision-making of PCa (44). Radiolysis methods have shown high predictive performance in the diagnosis of benign and malignant PCa based on GS scores (45). Machine learning is the core of artificial intelligence, which mainly studies how computers can build models from data and use the established models to make predictions on new inputs. At present, the machine-learning algorithms widely used in the medical field include ANN, SVM, KNN, DT, RF, and LR. Chen studied the multi-gene model constructed by machine learning to predict prostate cancer and concluded that the accuracy of the RF model to identify prostate cancer was 94% with AUC = 0.94. Wang conducted a machine-learning prediction study of prostate cancer, using transmittal ultrasound video clips, and found that the AUCs of the SVM model in the validation set and test set were 0.78 and 0.75, respectively, and the diagnostic efficiency of the SVM model was higher than that of MRA-based diagnosis (AUC was 0.78 vs. 0.65/0.75 and 0.75 vs. 0.65/0.72, respectively). Li found that the MRI-based SVM model had high diagnostic efficiency and stability (46). In this study, the accuracy of the SVM model for the diagnosis of clinically significant prostate cancer was 95%, and the AUC was 0.845, which is similar to the effect in Li’s study. The MRI radiomics model based on T2WI and ADC constructed by Li can improve the diagnosis of csPCa, and its efficacy in the validation set (AUC = 0.98) (47) is higher than that of the ANN model in this study (AUC = 0.855). This outcome may be due to the inclusion of ADC and clinical risk factors in Li’s study. Chen also constructed a radiomics model and found that the radiomics model showed higher predictive performance than a PI-RADS v2 evaluation in the identification and invasiveness assessment of PCa. In the identification of PCa, the AUC of the model in the validation set was 0.999, and the AUC of the invasiveness assessment was 0.93 (48), which were higher than those in this study. In addition, Nketiah combined the texture features extracted from T2WI, quantitative ADC values, and DCE pharmacokinetics parameters to predict the GS grade of PCa, and its prediction performance (AUC = 0.91) was higher than that of the ANN model in this study (AUC = 0.855). This outcome may be due to the small number of patients included (23 cases) and the lack of an independent validation set (49).

In this study, six machine-learning models for the detection of prostate cancer were established based on the significant prostate cancer risk factors identified by single logistic analysis. The prostate cancer machine-learning model was comprehensively evaluated based on the model evaluation indexes. The results showed that the machine-learning models could effectively detect CS-PCa categories, and the ANN model was the best (sensitivity was 0.80%, accuracy was 88.6%, and AUC was 0.855). Porter studied six widely used prediction models and found that ANN was better than the other models in predicting the outcome of a prostate biopsy. Furthermore, when more input variables are used before the biopsy, many unnecessary biopsies can be avoided. This is consistent with the results obtained in this study (50). Nevertheless, the current study still has some shortcomings. First, this study was conducted at a single institution, so the number of cases was small. Subsequently, the sample was expanded and combined with multi-center clinical studies to verify the results. Second, this study only included prostate cancer cases and only evaluated the aggressiveness of prostate cancer. In the future, a study on the differential prediction of benign and malignant prostate cancer may be conducted. Third, the pathology of the prostatic biopsy was used as the control rather than the pathology of a radical prostatectomy, which may lead to bias (51). The final multimodal ultrasound test data are subjective to a certain extent. In particular, ultrasonic elasticity imaging relies on manual pressure, and the pressure intensity and frequency of different physicians differ, which may cause certain deviations in the elasticity scores.

In our study, we innovatively established machine learning to predict CS-PCa using PSA-related risk factors. The predictive model based on machine learning has superior diagnostic efficiency in CS-PCa. In the ANN model, the sensitivity, PPV and NPV of the model are higher than the other five machine learning models. So machine learning models based on psa related risk factors can help radiologists improve their diagnosis. In future work, we intend to combine various machine learning models and deep learning models in the application of CS-PCa ultrasound diagnosis to improve the diagnosis of significant prostate cancer. In conclusion, establishing a machine learning model using CS-PCa risk factor groups can be an important method for detecting PCa. Multimodal ultrasound combined with PSA-related indicators has high clinical value in the diagnosis of CS-PCa.

The original contributions presented in the study are included in the article/Supplementary Materials, further inquiries can be directed to the corresponding author/s.

The studies involving human participants were reviewed and approved by the Ethics Committee of Dongyang People’s Hospital. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements. Written informed consent was not obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.

Conceptualization, MZ, YL, KW and ZW; methodology, MZ, YL, KW and ZW; validation, MZ and YL; formal analysis, JY and LC; investigation, ZW and YD; resources, MZ and ZW; data curation, KW, JT, ZH and YJ; writing—original draft preparation, MZ and YL; writing—review and editing, YD and XS; visualization, MZ and YL; supervision, ZW and LC; project administration, MZ and ZW; funding acquisition, ZW and LC. All authors have read and agreed to the published version of the manuscript.