Currently, tissue biopsy remains the gold standard for PCa diagnosis. Pathologic biopsy is also a common method for estimating PCa aggressiveness, but it has various complications limiting its clinical use. Furthermore, this evaluation typically relies on a solitary biopsy of a potentially heterogeneous tumor, which can merely capture a snapshot of its biological characteristics [12]. In the present dual-center retrospective study, we built and tested radiomics models for predicting PCa aggressiveness regarding GS and positive needles of systematic biopsy. The models were validated using internal and external validation sets. The established radiomics models provided a quantifiable and individualized tool for predicting PCa aggressiveness, thus helping in clinical decision-making.

Serum PSA is a specific marker of PCa and the only tumor marker with organ specificity. In this study, TPSA, FPSA, FPSA/TPSA, and PSAD in the GS ≤ 7 and GS > 7 groups differed considerably. It may be due to the biological differences between tumors with different grades. With the increase of serum PSA related indicators, GS will also increase. Moreover, TPSA, FPSA, FPSA/TPSA, and PSAD in the positive needles ≥ 6 group were higher than those in the positive needles < 6 group. We guessed that the number of positive needles is an indicator of the extent and localization of cancerous tissue of PCa. A higher number of positive needles (≥ 6) suggests a more extensive disease involvement.

Patient management for PCa requires an accurate evaluation of potential tumor aggressiveness [13, 14]. GS is a histopathological grading system used to assess the aggressiveness of PCa. Tumors with a higher GS are typically more aggressive and have a poorer prognosis. Previous studies investigated the radiomics methods to identify GS for estimating PCa aggressiveness. Gong’s group [15] presented a biparametric MRI radiomics for discriminating between patients with GS ≤ 7 and those with GS > 7, achieving satisfactory performance, with AUCs of 0.811 and 0.788 in the training and test cohorts, respectively. In another study, a multiparametric MRI-based radiomics signature demonstrated the potential to noninvasively distinguish between indolent and aggressive PCa [16]. However, a few studies mainly predicted GS, paying little attention to positive needles. Besides, they were single-center studies without external validation, whose results might be less robust and generalizable. The present study was a 2-center clinical study. We predicted simultaneously GS and positive needles of systematic biopsy based on whole prostate gland segmentation. This method may aid clinical doctors in evaluating comprehensively PCa aggressiveness, thus enabling personalized medicine. Highly aggressive PCa progresses rapidly and requires early intervention such as radical surgery or radiation therapy.

This study employed a radiomics analysis based on the whole prostate gland. It developed a noninvasive method for predicting GS ≤ 7 and GS > 7, and positive needles ≥ 6 and positive needles < 6, among patients with PCa. In agreement with previous studies [17, 18], LR was chosen as the classifier to construct radiomics models in this study, suggesting its advantages in assessing PCa aggressiveness. One of the reasons might be that LR was particularly effective in binary classification tasks [19]. For GS prediction, the radiomics models demonstrated a moderate-to-good diagnostic performance with an AUC of 0.811 [95% confidence interval (CI), 0.73–0.90] in the training sets, 0.814 (95% CI 0.69–0.93) in the internal validation sets, and 0.717 (95% CI 0.57–0.86) in the external validation sets. For positive needle prediction, the radiomics models demonstrated satisfactory predictive efficiency with an AUC of 0.806 (95% CI 0.71–0.89) in the training sets, 0.811 (95% CI 0.69–0.93) in the internal validation sets, and 0.791 (95% CI 0.65–0.93) in the external validation sets. Considering the external validation sets, the radiomics models performed satisfactorily in identifying positive needles compared with identifying GS. The reason might be that positive needles included more lesions and invaded regions with PCa. Specifically, 11 features were chosen for the GS prediction, whereas 5 features were chosen for positive needle prediction. Among all the features, the texture and wavelet features were vital, providing more information regarding tumor heterogeneity, which has been confirmed in several other studies [20,21,22]. We further analyzed the radiomics features and observed that Gradient_glszm_LargeAreaHighGrayLevelEmphasis and wavelet-LLH_firstorder_Kurtosis were common features of GS and positive needle prediction, although they did not contribute the most. The texture features can serve as a biomarker for predicting the presence of clinically remarkable PCa [23]. The wavelet filter disassembles the original images in various directions and reveals multidimensional spatial heterogeneity, which can assist in revealing tumor heterogeneity that may not be detectable in the original images [24, 25]. Therefore, we concluded that the texture and wavelet features might be the most helpful in predicting GS and positive needles of systematic biopsy to estimate PCa aggressiveness. Clinicians might be alerted to a potentially highly aggressive PCa using radiomics as a noninvasive method in our workflow.

Numerous studies have suggested the wide use of MRI for the diagnosis, staging, and treatment monitoring of various tumor types [26,27,28]. In this study, we initially predicted the PCa aggressiveness from sFOV HR-T2WI and post-contrast delayed scan sequences. The use of only these two sequences may limit the diversity of radiomic features that can be extracted. Different imaging modalities and sequences capture different aspects of tissue properties, and a more comprehensive approach incorporating multiple sequences and modalities could provide a richer set of features for analysis. However, the features that established radiomics models were all derived from the VOI on sFOV HR-T2WI series, indicating the crucial role of sFOV HR-T2WI images in providing aggressiveness-relevant information. The heavily weighted radiomic features from sFOV HR-T2WI series may effectively reflect more potential morphological and heterogeneity features of tumors, with higher spatial resolution and contrast. The result of this study was in line with that of previous studies [29, 30]. T2WI was considered more valuable than contrast-enhanced scanning sequences in reflecting tumor heterogeneity. Hence, choosing a valuable sequence is vital, avoiding time-consuming and laborious image segmentation. Therefore, we concluded that radiomics based on sFOV HR-T2WI might contribute to assessing PCa aggressiveness and risk stratification without additional MRI sequences such as dynamic contrast-enhanced MRI.

A large number of studies confirmed that tumor heterogeneity was not only solely determined by the tumor itself but also closely related to TME, which was perceived as a major determinant of cancer progression and aggressiveness [31, 32]. In this study, the prostate gland segmentation included lesion and TME information, providing a more comprehensive description of tumor-related information, which was helpful for tumor diagnosis and prognosis assessment. Furthermore, according to previous studies, approximately 50% of the radiomics features for prostate lesion segmentation were unstable, whereas only 20% of the radiomics features for gland segmentation were unstable [3]. In this study, only 1% of radiomics features from the gland had an ICC value < 0.8, indicating that the radiomics features from the whole prostate exhibited improved reproducibility and stability. This also highlighted the possibility of achieving a fully automatic segmentation of the whole prostate gland and promoting the noninvasive prediction of PCa aggressiveness in the future [33].

Data balancing tools, such as upsampling, downsampling, or SMOTE, have been demonstrated to considerably improve the predictive performance of radiomics models. In this study, the performance of radiomics models distinctly improved after data upsampling or SMOTE, which was in an excellent agreement with previous findings [3, 34].

This study had several limitations. First, it was a retrospective study, and although external validation data was included, the total sample size collected was small. In future studies, we plan to increase the sample size and conduct more external validations from multiple centers to obtain higher levels of clinical evidence. Second, only sFOV HR-T2WI and post-contrast delayed scan sequences were used. Other sequences, such as DWI, ADC, and dynamic contrast-enhanced MRI, could be worth exploring. Third, the radiomics models based on lesion segmentation was not involved, which will be investigated in our future works. Furthermore, the pathologic evaluations of 2 hospitals in this study did not guarantee that the results were obtained by the same urology specialist pathologist at the same time, introducing a certain degree of subjective variability.