2.1. Detection and Localization of PCa

McNeal in 1988 divided the prostate gland into a peripheral zone (PZ), central zone and transitional zone (TZ), consisting of 70%, 25% and 5% of the prostate volume, respectively [32]. PCa arises in the PZ in 80% of cases; less commonly it originates in the TZ or, rarely, in the central zone. Some studies have found that tumors originating in the peripheral zone are the most aggressive [33].

T1W, T2W, DWI and DCE images provide important anatomical and functional information; however, about 25% of PCa in TZ may either not be recognized or mistaken for a benign prostatic hyperplasia nodule by mpMRI [34]

The use of computer-aided diagnosis tools used to complement radiologists’ assessments increases sensitivity and specificity in detecting PCa [35]. Recent radiomic publications for detection and location of PCa are summarized in Table 1.

Recently, there has been considerable interest in the role of texture features extracted from the computer (or radiomics) for detection [34]. In literature, there are many articles that aim to provide a radiomics approach based on mpMRI for the detection and localization of PCa in the peripheral and transition zones [36,37].

In 2016, Cameron et al. [38] proposed a quantitative comprehensive feature model called MAPS (morphology, asymmetry, physiology and size) based on radiomics for automatic detection of PCa and achieved an accuracy, sensitivity and specificity of 87%, 86% and 88%, respectively.

In 2018, Khalvati et al. [39] proposed an optimal multiscale radiomics-driven framework (MPCaD) for automated localization and detection of PCa. MPCaD leverages the full set of voxel-level quantitative radiomic features and incorporates region-level feature descriptors in a pipeline to better characterize and detect tumor regions. This framework incorporating computed high-b DWI (CHBDWI) and correlated diffusion imaging and was evaluated using clinical prostate mpMRI data from 30 patients. The authors demonstrated that the proposed framework exhibits enhanced differentiation of tumor and healthy tissue, reaching, with the full modalities model, a sensitivity, specificity and accuracy of 0.82, 0.89, and 0.86, respectively. The authors confirmed that the quantitative radiomic characteristics extracted from magnetic resonance imaging of the prostate can be used to detect and localize PCa.

Another study by Wibmer et al. [40] investigated whether Haralick texture analysis [41] of prostate MRI was useful for PCa detection on both the peripheral and transitional zone. They enrolled 147 patients extracted texture features (energy, entropy, correlation, homogeneity and inertia) from T2-weighted images and ADC maps. They found that in peripheral zone, on both T2-weighted images and ADC maps, entropy and inertia were significantly higher in prostate cancer areas than in noncancerous areas, whereas energy, correlation and homogeneity areas were significantly lower. For the transitional zone on ADC maps, entropy and inertia were significantly higher in the cancer areas than in noncancerous areas, whereas energy, correlation and homogeneity were significantly lower; on T2-weighted images, inertia was significantly higher in cancer lesions than in noncancerous areas whereas correlation was significantly lower. They concluded that Haralick-based texture features allowed for differentiating benign and malignant prostate tissue.

Nketiah et al. [42] performed a single arm, multicenter study to evaluate the potential of T2-weighted image-derived textural features for quantitative assessment of peripheral zone prostate cancer aggressiveness. They extracted traditional intensity histogram features from T2-weighted images and ADC maps and second- and high-order statistical image textural features based on gray level co-occurrence matrix (GLCM) and gray level run length matrix (GLRLM) from T2-weighted images. Spearman correlations were used to evaluate association between textural features and PCa grade groups. Mann–Whitney U-tests and support vector machine (SVM) classifier were evaluated to differentiate and classify low-(grade group 1) vs. intermediate/high-(grade group ≥ 2) aggressive cancers, respectively. The cross-validation scheme employed in support vector machine classifier training and testing across six institutional centers, and it works in the following way: at each iteration, data from one institution was held out for testing, and data from the remaining five institutions used for training. The mean classification accuracy across the centers was highest for the combined ADC and T2W intensity-textural features (84%) compared to ADC histogram (75%), T2W histogram (72%), T2W textural (72%) features alone or T2W histogram and texture (77%), T2W and ADC histogram (79%) combined.

2.2. Application of MR-Derived Metrics in PCa

The main application of mpMRI was localization and staging of PCa. Radiomics applied to mpMRI of the prostate have a wide application field, including tumor localization and detection, prediction of prognosis and follow-up after treatment. The great amount of data generation and the increasing volume of imaging data demanded the application of computerized methods to analyze mpMRI data and extract useful information.

Computerized quantitative analysis may lead to effective, accurate and reproducible analysis of large amounts of data from mpMRI. The main advantage of quantitative MRI will be elimination of subjective assessment by radiologists. A way to improve the characterization of focal lesions at multiparametric MR imaging could be to use computer-aided diagnosis (CAD) systems. CAD and artificial intelligence tools have been investigated for PCa diagnosis with mpMRI data. Several studies have been found CAD effective in aiding radiologists in PCa diagnosis. For example, Hambrock et al. [35] showed that the use of a CAD system in clinical condition could significantly improve the characterization of prostate lesions by less experienced readers. Niaf et al. [43] showed that a CAD system may improve the characterization of prostate lesions with mpMR imaging by increasing reading specificity. However, although numerous studies have shown promising results, the low specificity and high false-positive rate of CAD continue to be a major problem [44,45], so that further multicenter studies with large populations are needed for validation.

Another way to reduce the subjective evaluation by radiologists is the automatic segmentation method by the machine and deep learning approach. In the biomedical imaging field, target delineation is routinely used as the first step in any automated disease diagnosis system to obtain quantitative parameters from biomedical images. Deep learning algorithms have been applied in automatic segmentation of the prostate gland [46,47] with potential benefit for patient management personalization. As a future perspective, the integration of a deep learning network in radiological PACS would lead to a rapid and precise procedure of segmentation of the prostate gland, thus reducing interuser variability.

2.3. Prediction of Gleason Score and PI-RADS

Gleason score (GS) is the current clinical gold standard for prognostication of PCa [48]. GS allows for the stratification of patients into different risk groups based on architectural alterations of prostate tissue based on biopsy or prostatectomy [49], allowing discrimination between clinically significant and not significant PCa, defined as GS < 4 + 3, or as the maximum length of the tumor nucleus < 6 mm [50]. Some articles have confirmed that radiomics can predict GS in vivo [48]. Recent radiomic publications for prediction of GS are summarized in Table 2.

Many studies have been conducted in order to discriminate between clinically significant and not significant PCa using MRI radiomic features. Fehr et al. [51] in 2015 conducted a study on 147 patients and found that a combination of T2w and ADC magnetic resonance characteristics of the Haralick plot were able to distinguish low GS from intermediate and high GS with 92% accuracy. In another study, Cuocolo et al. [52] evaluated whether radiomic shape features derived from MR images could be effective in clinically significant PCa detection. They enrolled 75 patients and extracted ten shape features both from axial T2-weighted and ADC maps images, after lesion tridimensional segmentation. Using multivariable analysis, the parameter defined as surface area to volume ratio extracted from ADC maps was the strongest independent predictor of clinically significant PCa with AUC of 0.78, specificity of 97% and sensitivity of 56%.

Studies conducted to discriminate between clinically significant and not significant PCa using MRI radiomic features were performed using different MRI scanners (1.5T or 3T) and extracting radiomic features from T2-weighted and ADC maps [40,51,53,54,55]. These studies confirmed that radiomics can predict GS and allows differentiation between low to intermediate-high risk GS.

Therefore, in addition to confirming the presence of disease in vivo, radiomics is able to distinguish between clinically significant and not clinically significant PCa, allowing the early identification of patients who could be better candidates for active surveillance than definitive therapy [48].

The PI-RADS version 2.1, approved by the American College of Radiology, stratifies prostate lesions into different PCa risk categories [56]. However, this classification system has some limitations, including the potential interindividual variability in lesion categorization by radiologists [57]. Furthermore, at least 20–30% of the lesions indicated as PI-RADS 3 prove to be malignant lesions and this causes confusion in the diagnostic and therapeutic management of the patients [58].

Recently, some studies have verified the possibility of using radiomics to attribute PI-RADS scores. Wang et al. [57] used a machine learning-based system (support vector machine (SVM) based on radial basis function (RBF) kernel) to analyze radiomic features extracted from T2-weighted and DWI. They found that when radiomic features were added, the diagnostic performance of PI-RADS was improved with an increase in sensitivity from 79% to 94.4% in PCa in the peripheral zone, and from 73.4% to 91.6% in PCa in the transitional zone. The authors confirmed that MR radiomics can help to improve the performance of PI-RADS in clinically significant PCa.

In addition, Giambelluca et al. [58], using texture analysis software (MaZda 4.6), extracted first-, second- and third-order radiomic features from 46 PI-RADS 3 lesions segmented on T2-weighted and ADC maps. The authors found nine and six independent texture parameters on T2-weighted images and ADC maps, respectively, that significantly correlate with the final histological results. This confirmed that radiomic features can help to distinguish between PCa and nontumor tissue among PI_RADS 3 lesions.

Hou et al. [59] in 2020 developed a model integrating data extracted from T2W, DWI and ADC maps images of 271 patients. This model achieved promising performance in improving diagnostic accuracy in PI-RADS 3 by allowing clinically significant PCa to be differentiated from indolent and normal cases.

2.4. Prediction of Extracapsular Extension

Radical prostatectomy is considered the preferred approach for patients with localized PCa [60]. Accurate preoperative staging is important because any extracapsular extension (ECE) of the tumor influences the clinical decision-making process [61,62]. Various predictive models have been analyzed to try to understand the risk of ECE before surgery. The Partin tables and Memorial Sloan–Kettering (MSK) preradical prostatectomy nomograms are two of the most used models [63].

Many studies have shown that the accuracy of mpMRI for ECE detection varies from 62% to 76% but decreases up to 30% in the case of apex tumors [63].

Currently, neither predictive models nor mpMRI are effective in detecting the real risk of ECE due to their intrinsic limitations [63].

In recent years, a few studies proposed a radiomics approach to evaluate ECE. Ma et al. [63] proposed radiomic signatures based on T2W images to predict the side-specific probability of ECE for patients with PCa. They proposed a radiomics signature incorporating 17 selected features and demonstrated favorable discrimination capabilities in both the training and validation datasets, along with good calibration performances.

Losnegard et al. [64], in 2020, tested how the MRI radiomics approach performs in combination with preoperative clinical variables and radiology MRI interpretation. Texture features were extracted from T2-weighted and ADC maps using a Matlab toolbox (Lloyds function), and quantitative DCE time-series features were extracted using an in-house tool. Then, they used the extracted feature in a supervised machine learning setting in order to obtain the probabilities for patients to have ECE. A logistic regression model to obtain probabilities of ECE was calculated from clinical variables (PSA level, TNM stage, Gleason grade and score, percentage of positive core) and combined with radiomic features. They found that features extracted from T2-weighted and ADC maps were the best radiomic features while nonadditional benefit was added from DCE features. Moreover, they reached a good performance of combined models, with an AUC of 0.79 in prediction of ECE. They concluded that MR radiomics may represent a valuable adjunct to conventional prediction models for ECE.

The proposed radiomics signature had superior diagnostic performance compared, for example, to the visual assessments of radiologists, particularly for apical tumors [63].

In addition, Xu et al. [65] proposed a radiomics model based on mpMRI able to differentiate ECE and non-ECE lesions in preoperative time. They extracted radiomic features from T2-weighted images, DWI, ADC maps and DCE images using the Python package Pyradiomics (version 2.2.0). They retained 30 features by the use of a maximum relevancy and minimum redundancy algorithm. Finally, by the use of the least absolute shrinkage and selection operator (LASSO) regression algorithm, the final radiomics model was built and integrated with the clinical model to build a combined nomogram. They found that the combined nomogram outperformed the clinical model in diagnosing ECE, but was comparable with using the radiomics model alone.

A major limitation of all the reported studies, however, is that results are not validated on large population or in a multicenter study, and needs validation to have a role in future clinical application.

2.5. Prediction of Biochemical Recurrence after Treatment (Surgery or Radiotherapy)

Biochemical recurrence (BCR) occurs in 50% of patients who underwent radical prostatectomy, especially those with high-risk characteristics such as locally advanced disease (T3-4), positive margins (R1) or high Gleason score [66].

Adjuvant radiotherapy includes BCR-free survival, but its impact on overall survival is controversial and also a major cause of side effects [67].

For these reasons, radiotherapy (RT) is often provided exclusively in patients with BCR [67]. MpMRI allows recognition of seminal vesicle invasion or ECE, both of which are important predictors of biochemical relapse-free survival at five years [68].

To date, there are only a few studies on radiomics prediction of BCR. Bourbonne et al. [69] trained and validated a radiomics-based model that appears to be predictive of BCR and biochemical recurrence-free survival after radical prostatectomy in patients with PCa. They extracted first-order intensity metrics, second- and third-order textural features from T2-weighted and ADC maps of 107 patients. Then, a univariate and multivariate Cox regression analysis was performed to identify independent factors and the correlation with BCR was assessed. This model has a negative predictive value of 96% and could be used to identify patients at very low risk of recurrence. It would also help stratify patients after radical prostatectomy by helping the clinician adapt postoperative management. Using this model, patients with very low risk of BCR could avoid RT, thus reducing side effects, while patients with high risk of BCR could be offered intensified postsurgical monitoring or adjuvant RT.

Zhong et al. [70], in a retrospective study conducted on 91 patients, evaluated the relationship between pretreatment mpMRI radiomic features and BCR in patients with localized prostate cancer. The authors of this study validated a model with an improved predictive value over conventional imaging metrics that could facilitate MRI-based prognostic predictions and assist in the decision making for the individual treatment.