Predictive role of T2WI and ADC-derived texture parameters in differentiating Gleason score 3 + 4 and 4 + 3 prostate cancer

Abstract

BACKGROUND:

Since Gleason score (GS) 4 + 3 prostate cancer (PCa) has a worse prognosis than GS 3 + 4 PCa, differentiating these two types of PCa is of clinical significance.

OBJECTIVE:

To assess the predictive roles of using T2WI and ADC-derived image texture parameters in differentiating GS 3 + 4 from GS 4 + 3 PCa.

METHODS:

Forty-eight PCa patients of GS 3 + 4 and 37 patients of GS 4 + 3 are retrieved and randomly divided into training (60%) and testing (40%) sets. Axial image showing the maximum tumor size is selected in the T2WI and ADC maps for further image texture feature analysis. Three hundred texture features are computed from each region of interest (ROI) using MaZda software. Feature reduction is implemented to obtain 30 optimal features, which are then used to generate the most discriminative features (MDF). Receiver operating characteristic (ROC) curve analysis is performed on MDF values in the training sets to achieve cutoff values for determining the correct rates of discrimination between two Gleason patterns in the testing sets.

RESULTS:

ROC analysis on T2WI and ADC-derived MDF values in the training set (n = 51) results in a mean area under the curve (AUC) of 0.953±0.025 (with sensitivity 0.9274±0.0615 and specificity 0.897±0.069), and 0.985±0.013 (with sensitivity 0.9636±0.0446 and specificity 0.9726±0.0258), respectively. Using the corresponding MDF cutoffs, 95.3% (ranges from 76.5% to 100%) and 94.1% (ranged from 76.5% to 100%) of test cases (n = 34) are correctly discriminated using T2WI and ADC-derived MDF values, respectively.

CONCLUSIONS:

The study demonstrates that using T2WI and ADC-derived image texture parameters has a potential predictive role in differentiating GS 3 + 4 and GS 4 + 3 PCa.

Keywords

Classification of prostate cancer magnetic resonance imaging gleason score image texture analysis

1 Introduction

GS is widely accepted as an important prognostic factor in all treatments for PCa. Designed by Dr. Donald F Gleason in the 1960s and 1970s, GS is based on the histologic pattern of cancer cell arrangement in H&E stained sections [1]. It is the sum of the primary and secondary modes, ranging from 2 to 10, with each mode ranging from1 (well-differentiated) to 5 (poor-differentiated) [2]. A higher GS most likely means rapid growth and spread of the cancer cells. Although the total score of GS 3 + 4 and GS 4 + 3 are both 7, their biology characteristics quite differed. GS 4 + 3 PCa has a higher hazard ratio (5.1) than GS 3 + 4 PCa (1.9) relative to GS 6 PCa [3], and its mortality increased three times compared with GS 3 + 4 PCa [4]. A systematic review also confirmed the heterogeneous prognosis of GS 7 cancers [5]. To reflect these differences, a new PCa histological grading system was proposed at the 2014 consensus meeting of the international society of urological pathology, in which GS = 7 PCa was divided into two groups, GS 3 + 4 and GS 4 + 3 [6].

PIRADS v1 [7] and v2 [8], and the evolving v2.1 [9] have been developed to assess the risk of lesions being PCa. Although the inter-reader agreement of PIRADS v1 and v2 has been reported as excellent in one study, 0.71 and 0.81, respectively, their diagnostic performance differed. Another study reported the inter-observer agreement in two readers was only 0.48 both in PIRADS v1 and v2 [10]. At the same time, it is difficult for conventional imaging techniques to analyze the huge image digital features caused by cell, physiological and genetic variations in the image, which likewise cannot be recognized by the human eye [11]. These promoted the generation of radiomics and artificial intelligence, which extract texture features from standard medical images to improve the accuracy of diagnosis and prediction [12]. Increasing evidence showed that these methods can be used to enhance non-invasive tumor characterization, including tumor detection [13], preoperative risk stratification [14], classification [15], prediction of certain tumor molecular features [16], association with tumor spread, better prediction of treatment response and prognosis [17]. It was reported that textural features extracted from ADC images showed the greatest clinical application potential in MRI of PCa [18], and lower angular second moment and higher entropy were seen in GS 4 + 3 PCa than GS 3 + 4 PCa [19, 20]. They are concrete partial texture features, we hypothesized that the most discriminant features (MDF) calculated from texture features could help differentiate PCa of GS 3 + 4 from GS 4 + 3, so the purpose of this study is to assess the predictive roles of T2WI and ADC-derived texture parameters in differentiating the two types of PCa.

2 Materials and methods

2.1 Patients selection

The Hospital Ethics Committee approved our study and waived informed consent. All patients’ data were analyzed anonymously. We searched for prostate cancer patients between June 2013 and May 2019. The inclusion criteria included: (1) Pathological examination of the resected or biopsy specimen showed prostate cancer; (2) GS 3 + 4 and GS 4 + 3 PCa; (3) multiparametric MRI (mpMRI) of the prostate, including at least T2WI and DWI, was performed before surgical resection or biopsy; (4) MRI showed lesions larger than 5 mm; (5) Surgical resection or biopsy was performed within 2 weeks after mpMRI; (6) MRI images have no artifacts and meet the post-processing requirements; (7) MR scanning was performed with Siemens 3T scanner. The exclusion criteria included: (1) mpMRI scans were performed after surgery or biopsy; (2) The images had artifacts; (3) Poor image quality or smaller than 5 mm; (4) Other GS PCa.

2.2 MRI protocol

Prostate mpMRI was performed in Siemens 3T MRI scanner (MAGNETOM Skyra, Siemens Healthcare, Erlangen, Germany) using an 18-channel body coil array. The MRI scan covered the entire pelvic. The scanning sequence included TW2I, DWI, with or without PWI. The T2WI scanning parameters without fat saturation were set as following: repeat time 6750 ms, echo time 104 ms, slice thickness 3 mm, slice spacing 3 mm, FOV 180×180 mm, matrix 346×384, flip angle 160 degrees, NEX = 2. The b values in DWI scan were 0, 50, 200, 400, 600, 800 and 1500 s/mm2. ADC images were automatically reconstructed for qualitative and quantitative evaluation. ADC images generated from b value of 800 s/mm2 were used for analysis.

2.3 Image segmentation and feature extraction

MR images were retrieved from the Picture Archiving and Communication System (PACS). The sections showing the maximum lesion size in axial T2WI and ADC were selected for further texture extraction. Images were imported to MaZda software. A radiologist with 5 years of experience in urological radiology who was blinded to patients’ clinical characteristics and GS draw the region of interest (ROI) manually, 1-2 mm from the edge of lesions, avoiding necrosis and adjacent non-neoplastic structures. For each patient, two ROIs were delineated on axial T2WI and ADC images (Fig. 1). A total of 170 ROIs from 85 patients were obtained in this study. Feature extraction was performed using MaZda software. The images were normalized in the range of μ-3SD and μ+3SD to minimize the influence of contrast and brightness variation. Six common feature groups including 300 radiomics parameters (histogram, gradient, run-length matrix, co-occurrence matrix, autoregressive model, and wavelet transform) were extracted from MaZda (Table 1).

Fig. 1

The delineation of ROI. ROI was delineated on the maximum section of the lesion, 1-2 mm from the edge of the lesion, avoiding adjacent structures on T2WI and ADC, respectively.

Table 1

Texture features extracted in the study

	Texture Features
Histogram	Mean, variance, skewness, kurtosis, and 1st, 10th, 50th, 90th, and 99th percentiles
Gradient	Mean, variance, skewness, kurtosis, and percentage of pixels with nonzero gradient
COM	Angular second moment, contrast, correlation, sum of squares, inverse difference moment, sum average, sum variance, sum entropy, entropy, difference variance, and difference entropy; parameters computed up to 20 times, for (d,0), (0,d), (d,d), (d,-d) where the distance d can take values of 1, 2, 3, 4, and 5
RLM	Run-length nonuniformity, gray-level nonuniformity, long-run emphasis, short-run emphasis, and fraction of image in runs; parameters computed for horizontal, 45°, vertical, and 135° directions
AM	θ1, θ2, θ3, θ4, σ
Wavelet transform	Energy of wavelet coefficients in low-frequency sub-bands, horizontal high-frequency sub-bands, vertical high-frequency sub-bands, and diagonal high-frequency sub-bands at successive scales

COM, co-occurrence matrix; RLM, run-length matrix; AM, autoregressive model.

2.4 Intra-observer agreement

To evaluate the intra-observer agreement, the same radiologist repeated the ROI delineation of 20 random samples at an interval of 2 months. The area of ROI was compared. Intraclass correlation coefficient (ICC) analysis was used to evaluate the intra-observer agreement. An ICC > 0.75 indicates satisfactory agreement.

2.5 Dimensionality reduction and radiomics feature selection

The texture features of each ROI are extracted by MaZda software, and then a combination of Fisher coefficients (a ratio of between-class to within-class variance), minimization of both classification error probability and average correlation coefficients (POE+ACC), and mutual information (MI) coefficients were used to screen out 30 optimal features that were most significant for the discrimination of two pathological types. Linear discriminant analysis (LDA), a built-in toolkit in B11 statistical software, was used to obtain the most discriminative features (MDF) of optimal texture features.

2.6 Training set model and validation of the testing set

Random number table method was used to select 60% samples as the training set (n = 51) and the remaining 40% samples serve as the test set (n = 34). AUC and diagnostic efficacy of MDF were obtained in the training set by using ROC analysis. The optimal cutoff of MDF was subsequently validated in the testing set, to calculate the correct classification rate. The 30 texture features in the testing set were consistent with those in the training set. To avoid overfitting, fivefold cross-validation was applied.

2.7 Statistical analysis

Continuous variables were expressed as mean±standard deviation (SD) or median and quartiles [M (25th percentile –75th percentile)] and Mann Whitney test was used to compare the differences between the two groups. Categorical data were expressed as numbers and variables were compared using the Chi-squared test. The above statistical analyses were conducted using SPSS (version 22.0; SPSS, Inc., Chicago, IL, USA). The discrimination metrics of established models in the training set, including AUC, sensitivity and specificity, cutoff were calculated using MdeCalc software. A two-sided P value < 0.05 was considered significant.

3 Results

3.1 Patient characteristics

A total of 48 patients with GS 3 + 4 and 37 patients with GS 4 + 3 were included. The clinical and radiological characteristics of the included cases are shown in Table 2. There was no statistical difference in age between the two groups (68.6±7.2 y vs 66.9±8 y, P = 0.3467). Total prostate-specific antigen was higher in GS 4 + 3 patients than in GS 3 + 4 (37.48 [19.67–243.8] ng/ml vs 23.94 [12.19–56.15] ng/ml, P = 0.0084). GS 3 + 4 PCa mainly distributed in the peripheral zone while GS 4 + 3 PCa distributed in the peripheral + central zone (20/48 vs 17/37, P = 0.009). Lesion size (2.40 [1.53–3.53] cm vs 3 [2.05–4.15] cm, P = 0.0663) and incidence of bone metastasis (3/48 vs 6/37, P = 0.1687) did not differ statistically.

Table 2
Clinical and radiological characteristics of the included cases with GS 3 + 4 and 4 + 3

Characteristics GS 3 + 4 (n = 48) GS 4 + 3 (n = 37) P value

Age (y) 68.6±7.2 66.9±8 0.3467

tPSA (ng/ml) 23.94 [12.19–56.15] 37.48 [19.67–243.8] 0.0084^*

Tumor location

Peripheral zone 20 7 Overall 0.0339^*

Central zone 17 13 P vs C 0.169

Peripheral+Central zone 11 17 P vs P + C 0.009^*

C vs P + C 0.186

Lesion size (cm) 2.40 [1.53–3.53] 3 [2.05–4.15] 0.0663

Bone metastasis 3/48 6/37 0.1687

Characteristics	GS 3 + 4 (n = 48)	GS 4 + 3 (n = 37)	P value
Age (y)	68.6±7.2	66.9±8	0.3467
tPSA (ng/ml)	23.94 [12.19–56.15]	37.48 [19.67–243.8]	0.0084^*
Tumor location
Peripheral zone	20	7	Overall 0.0339^*
Central zone	17	13	P vs C 0.169
Peripheral+Central zone	11	17	P vs P + C 0.009^*
C vs P + C 0.186
Lesion size (cm)	2.40 [1.53–3.53]	3 [2.05–4.15]	0.0663
Bone metastasis	3/48	6/37	0.1687

The lesion size was defined as the maximum diameter on the cross-sectional image; tPSA, total prostate-specific antigen.

3.2 Diagnostic performance in the training set and validation in the testing set

Twenty tumors (10 GS 3 + 4 and 10 GS 4 + 3 lesions) were segmented by the same radiologist at an interval of 2 months, the areas of the lesions were compared, showing a strong correlation between the two measurements with Pearson correlation of 0.964 and ICC of 0.956.

Three hundred texture features were extracted from T2WI, ADC images, and 30 optimal radiomics features for each ROI were selected by reducing dimensionality through a combination of Fisher, POE+ACC, and MI coefficients.

For T2WI-derived texture features, MDFs in the training set were obtained through LDA in B11 analysis, showing a significant difference between the two groups (P < 0.0001). The average sensitivity and specificity of the five cross-validation tests were 0.9274±0.0615, 0.897±0.069, and the mean AUC was 0.953±0.025 (ranged from 0.915 [0.803–0.975] to 0.986 [0.905–1]). When the corresponding cutoff of MDF (ranged from –0.002 to 0.0033) was applied in the testing set, the average correct rate of discrimination was 95.3% (ranged from 76.5% to 100%). The training results with the median AUC and corresponding testing results are shown in Fig. 2.

Fig. 2

Diagnostic performance of MDF derived from T2WI in the training set and validation in the testing set (the median AUC of the training set in the 5-fold cross-validation). 11/51 lesions were misclassified by MDF in the training set (a); the diagnostic AUC was 0.948, sensitivity was 86.4%, specificity was 93.1% (b); Using the cut-off MDF of –0.0002, 8/34 lesions in the testing set were misclassified (c).

For ADC-derived texture features, MDFs in the training set obtained through LDA shows a significant difference between the two groups (P < 0.0001). The average sensitivity and specificity of the five cross-validation tests were 0.9636±0.0446, 0.9726±0.0258, and the mean AUC was 0.985±0.013 (ranged from 0.967 [0.875–0.997] to 1 [0.930–1]). When the corresponding cutoff of MDF (ranged from –0.0041 to 0.005) was applied in the testing set, the average correct rate of discrimination was 94.1% (ranged from 76.5 to 100%). The training results with the median AUC and corresponding testing results are shown in Fig. 3.

Fig. 3

Diagnostic performance of MDF derived from ADC in the training set and validation in the testing set (the median AUC of the training set in the fivefold cross-validation). 6/51 lesions were misclassified by MDF in the training set (a); the diagnostic AUC was 0.983, sensitivity was 90.9%, specificity was 100% (b); Using the cut-off MDF of 0.0016, all 34 lesions in the testing set were classified correctly (c).

4 Discussion

This study showed that T2WI and ADC-derived texture features could help distinguish PCa of GS 3 + 4 from GS 4 + 3. The radiomics model based on MDF derived from T2WI and ADC texture features could correctly distinguish PCa between GS 3 + 4 and GS 4 + 3 by 95.3% and 94.1%.

The spatial and morphological heterogeneity of PCa affects its tumor grading and classification [21]. As PCa histology, the developed GS can help predict the aggressiveness of PCa [22] and plays an important role in PCa management strategies [23]. The biological characteristics of PCa with a GS 7 were heterogeneous, with a 40% risk of PCa progression within 5 years for GS 4 + 3 and only 15% for GS 3 + 4 [24]. A non-invasive distinction between GS 3 + 4 and 4 + 3 would benefit PCa patients. Mp-MRI is considered to be the most sensitive and specific imaging tool for the detection, characterization and staging of PCa. ROC curve analysis showed that the AUC of mp-MRI to distinguish GS 6 from GS≥7 (3 + 4) PCa was 0.73, and the discrimination of GS≤7 (3 + 4) from GS≥7 (4 + 3) is as high as 0.9 [25]. However, the AUC for the diagnosis of GS 4 + 3 PCa was only 0.65, and the intra-observer consistency was moderate [26]. Although mpMRI has been widely used in the assessment of prostate cancer, it can hardly differentiate GS 3 + 4 and 4 + 3 visually. Using quantitative imaging features of mpMRI and coding intratumoral heterogeneity to predict GS as a non-invasive biomarker for PCa is attracting more and more attention [27]. A meta data analysis indicated a higher pooled ADC value 0.91 × 10^–3 mm²/s of GS 3 + 4 PCa than GS 4 + 3 PCa (0.80 × 10^–3 mm²/s), however, the ADC values of the types of PCa overlapped [28], another systematic review showed ADC of peripheral PCa correlated moderately with GS and ADC of transitional PCa correlated weakly with GS [29]. Emerging textural analysis and artificial intelligence may help identify the two types of PCa. In this study, we extracted the T2WI and ADC radiologic features of GS 3 + 4 and GS 4 + 3 PCa and explored the potential value of the invisible radiologic features to distinguish the two types of tumors. We found that MDF derived from T2WI and ADC images had a comparative performance in the discrimination of GS 3 + 4 and GS 4 + 3 PCa.

Some studies have shown that the angular second moment and entropy derived from T2WI are related to GS. The angular second moment of GS 4 + 3 PCa was lower than that of GS 3 + 4, while the entropy was higher than GS 3 + 4, which can be used as a potential diagnostic marker sensitive to pathological grading [30], other studies have shown that the entropy and energy of ADC maps can help distinguish PCa with different GS [31]. MR texture features were proven to be related to tumor histopathology [32]. In this study, we selected the best 30 texture features, most of them were second-order texture features, and obtained the MDFs through LDA analysis, both T2WI- and ADC-derived MDFs revealed a good performance in distinguishing GS 3 + 4 from GS 4 + 3 PCa. Our study was similar to previous studies, in which T2WI images could provide the most important features in predicting GS 3 + 4 and 4 + 3 although AUC value of ADC texture features was higher with 62.09% than T2WI with 54.92% [27].

The clinical characteristics of GS 3 + 4 and GS 4 + 3 PCa also have certain differential significance, in this study, GS 4 + 3 PCa has a higher total prostate-specific antigen, mainly distributed in peripheral+central areas, different from GS 3 + 4 PCa which mainly distributed in the peripheral zone. Given the different pathological characteristics and clinical prognosis of the two kinds of tumors, the correct differentiation of the two kinds of tumors can help clinicians to treat them aggressively when necessary, but at the same time avoid overtreatment. The combination of clinical features and texture features may help us to distinguish two kinds of PCa better.

Our research also has some limitations, first, it was a retrospective study with a small sample size. Second, we did not test it in another patient sample with a different MRI scanner. Third, only T2WI and ADC were used in the analysis, perfusion-weighted imaging with more images and dynamic enhancement parameters may potentially improve the performance metrics for predicting GS. Future research can focus on integrating texture analysis and more advanced machine learning techniques to better explain the role of texture analysis in PCa clinical practice.

In conclusion, we in this study demonstrated that T2WI and ADC derived image texture features can be used to distinguish GS 3 + 4 from GS 4 + 3 PCa and achieve a comparative performance in discrimination.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Funding

This research was supported by the National Natural Science Foundation of China (Grant numbers: 81171307, 81671656).

References

Humphrey

P.A.

, Gleason grading and prognostic factors in carcinoma of the prostate, Mod Pathol 17 (2004), 292–306.

Gleason

D.F.

and Mellinger

G.T.

, Prediction of prognosis for prostatic adenocarcinoma by combined histological grading and clinical staging, J Urol 111 (1974), 58–64.

Epstein

J.I.

, Zelefsky

M.J.

, Sjoberg

D.D.

, et al., A Contemporary Prostate Cancer Grading System: A Validated Alternative to the Gleason Score, Eur Urol 69 (2016), 428–435.

Stark

J.R.

, Perner

, Stampfer

M.J.

, et al., Gleason score and lethal prostate cancer: does 3+4=4+3? J Clin Oncol 27 (2009), 3459–3464.

Kane

C.J.

, Eggener

S.E.

, Shindel

A.W.

and Andriole

G.L.

, Variability in Outcomes for Patients with Intermediate-risk Prostate Cancer (Gleason Score 7, International Society of Urological Pathology Gleason Group 2-3) and Implications for Risk Stratification: A Systematic Review, Eur Urol Focus 3 (2017), 487–497.

Epstein

J.I.

, Egevad

, Amin

M.B.

, et al., The 2014 International Society of Urological Pathology (ISUP) Consensus Conference on Gleason Grading of Prostatic Carcinoma: Definition of Grading Patterns and Proposal for a New Grading System, Am J Surg Pathol 40 (2016), 244–252.

Barentsz

J.O.

, Richenberg

, Clements

, et al., European Society of Urogenital, ESUR prostate MR guidelines 2012, Eur Radiol 22 (2012), 746–757.

Barrett

, Turkbey

and Choyke

P.L.

, PI-RADS version 2: what you need to know, Clinical Radiology 70 (2015), 1165–1176.

Latifoltojar

, Appayya

M.B.

, Barrett

and Punwani

, Similarities and differences between Likert and PIRADS v2.1 scores of prostate multiparametric MRI: a pictorial review of histology-validated cases, Clinical Radiology 74 (2019), 895.e1–895.e15.

10.

Schaudinn

, Gawlitza

, Mucha

, et al., Comparison of PI-RADS v1 and v2 for multiparametric MRI detection of prostate cancer with whole-mount histological workup as reference standard, European Journal of Radiology 116 (2019), 180–185.

11.

Jiang

, Jiang

, Zhi

, et al., Artificial intelligence in healthcare: past, present and future, Stroke Vasc Neurol 2 (2017), 230–243.

12.

Lambin

, Leijenaar

R.T.H.

, Deist

T.M.

, et al., Radiomics: the bridge between medical imaging and personalized medicine, Nat Rev Clin Oncol 14 (2017), 749–762.

13.

Becker

A.S.

, Marcon

, Ghafoor

, et al., Deep Learning in Mammography: Diagnostic Accuracy of a Multipurpose Image Analysis Software in the Detection of Breast Cancer, Invest Radiol 52 (2017), 434–440.

14.

Ueno

, Forghani

, et al., Endometrial Carcinoma: MR Imaging-based Texture Model for Preoperative Risk Stratification-A Preliminary Analysis, Radiology 284 (2017), 748–757.

15.

Al Ajmi

, Forghani

, Reinhold

, et al., Spectral multi-energy CT texture analysis with machine learning for tissue classification: an investigation using classification of benign parotid tumors as a testing paradigm, Eur Radiol 28 (2018), 2604–2611.

16.

Ganeshan

, Panayiotou

, Burnand

, et al., Tumor heterogeneity in non-small cell lung carcinoma assessed by CT texture analysis: a potential marker of survival, Eur Radiol 22 (2012), 796–802.

17.

Vallieres

, Freeman

C.R.

, Skamene

S.R.

and El Naqa

, A radiomics model from joint FDG-PET and MRI texture features for the prediction of lung metastases in soft-tissue sarcomas of the extremities, Phys Med Biol 60 (2015), 5471–5496.

18.

Patel

, Henry

and Scarsbrook

, The value of MR textural analysis in prostate cancer, Clin Radiol 74 (2019), 876–885.

19.

Pan

, Yang

, Shu

, et al., Application of texture analysis based on T2-weighted magnetic resonance images in discriminating Gleason scores of prostate cancer, J Xray Sci Technol 28 (2020), 1207–1218.

20.

Nketiah

, Elschot

, Kim

, et al., T2-weighted MRI-derived textural features reflect prostate cancer aggressiveness: preliminary results, Eur Radiol 27 (2017), 3050–3059.

21.

Tolkach

and Kristiansen

, The Heterogeneity of Prostate Cancer: A Practical Approach, Pathobiology 85 (2018), 108–116.

22.

Knuchel

, Gleason Score 6 - Prostate Cancer or Benign Variant? Oncol Res Treat 38 (2015), 629–632.

23.

Sehn

J.K.

, Prostate Cancer Pathology: Recent Updates and Controversies, Mo Med 115 (2018), 151–155.

24.

Chan

T.Y.

, Partin

A.W.

, Walsh

P.C.

and Epstein

J.I.

, Prognostic significance of Gleason score 3+4 versus Gleason score 4+3 tumor at radical prostatectomy, Urology 56 (2000), 823–827.

25.

Boesen

, Multiparametric MRI in detection and staging of prostate cancer, Dan Med J 64 (2017), B5327.

26.

Lim

C.S.

, McInnes

M.D.F.

, Flood

T.A.

, et al., Prostate iImaging Reporting and Data System, Version 2, Assessment categories and pathologic outcomes in patients with Gleason score 3+4=7 prostate cancer diagnosed at biopsy, AJR Am J Roentgenol 208 (2017), 1037–1044.

27.

Chaddad

, Niazi

, Probst

, et al., Predicting Gleason score of prostate cancer patients using radiomic analysis, Frontiers in Oncology 8 (2018), 630.

28.

Meyer

H.J.

, Wienke

and Surov

, Discrimination between clinical significant and insignificant prostate cancer with apparent diffusion coefficient - a systematic review and meta analysis, BMC Cancer 20 (2020), 482.

29.

Surov

, Meyer

H.J.

and Wienke

, Correlations between apparent diffusion coefficient and Gleason score in prostate cancer: A systematic review, Eur Urol Oncol 3 (2020), 489–497.