Biparametric magnetic resonance imaging-based radiomics model can improve the detection of dense and sparse prostate cancers

Introduction

Prostate cancer (PCa) is the second most common malignancy among males worldwide and is one of the leading causes of death. Multiparametric magnetic resonance imaging (mpMRI) plays an important role in the early detection of PCa. In recent years, studies have shown that simplified biparametric magnetic resonance imaging (bpMRI), which consists of only T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI), has comparable diagnostic efficacy and sensitivity to mpMRI, with good consistency among readers and shorter scanning time (1-3), demonstrating its advantages in clinical application. PCa tissues are heterogeneous; the tumor lesions can be dense foci composed of a high proportion of glands or sparse foci formed by malignant glands scattered within normal tissues. Differences in the growth patterns of these lesions will affect the detection of PCa on magnetic resonance imaging (MRI); specifically, the detection rate of sparse PCa is far lower than that of dense PCa. Langer et al. (4) classified peripheral zone (PZ) PCa lesions into sparse and dense types based on the difference in the proportion of normal glands and stroma and compared the signal characteristics of both with those of noncancerous tissues. The results showed that dense lesions had significantly lower T2WI and apparent diffusion coefficient (ADC) values than noncancer tissues, but no significant differences were observed between sparse lesions and noncancer tissues. van Houdt et al. (5) also found a similar pattern, in which the T2 and ADC values in sparse lesions were higher than those in dense lesions. Sparse lesions may not display well on conventional MRI.

In recent years, prostate-specific membrane antigen (PSMA)-targeting tracers have been increasingly used for evaluating PCa. Combining PSMA imaging with MRI can improve the assessment of local staging of PCa and optimize treatment planning (6). High PSMA uptake in primary lesions is associated with a higher Gleason score (GS) and poorer prognosis (7). However, some studies have shown that approximately 10% of PCa lesions do not uptake PSMA (8,9), and these lesions exhibit invasive growth patterns rather than dilated growth patterns (10). This observation led to an inference that the growth pattern of PCa may also affect the performance of diffusion sequences on MRI and potentially lead to false negative results (11).

With the introduction and development of radiomics, the extraction of effective radiomics features has been shown to help better describe the biological behavior of lesions in different environments (12). Xu et al. (13) used a bpMRI radiomics model to distinguish benign and malignant prostate tissues. The areas under the curve (AUCs) of the T2WI-, ADC-, and DWI-based models were 0.812, 0.893, and 0.775, respectively, and the overall AUC of the bpMRI model was 0.920. Prata et al. (14) had similar findings that the diagnostic efficacy of the fusion model of bpMRI and clinical features was higher than that of the ADC model, T2WI model, and clinical features model. The application of radiomics has improved the detection rate of PCa. However, there has been no radiomics-related research focused on sparse lesions.

The false negative results of MRI can directly lead to missed diagnoses for PCa. Therefore, this study attempted to use radiomics based on bpMRI combined with whole-mount pathology sections to analyze the differences in radiomics characteristics between infiltrative-growth (sparse) PCa lesions, dilated-growth (dense) PCa lesions, and noncancerous tissues. We present this article in accordance with the CLEAR reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2912/rc).

Methods

Study population

This retrospective study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Review Board of Fudan University Shanghai Cancer Center (No. 2005217-Exp2), and the requirement for individual consent for this analysis was waived due to the retrospective nature. From June 2020 to December 2022, patients from center 1 (Fudan University Shanghai Cancer Center) and center 2 (Zhongshan Hospital, Fudan University) with suspicious signs, such as palpable prostate nodules and/or prostate-specific antigen (PSA) levels >4 ng/mL, who underwent MRI examinations were searched. The inclusion criteria were as follows: (I) PCa confirmed by transrectal ultrasound-guided systematic biopsy with or without targeted biopsy after MRI examination, where the interval between biopsy and MRI was less than 1 month; (II) radical prostatectomy (RP) with an interval of less than 1 month between RP and MRI examination; (III) no systemic or focal treatment before RP; and (IV) complete information on histopathological whole-mount sections. The sparse lesions, dense lesions, and noncancerous tissues from center 1 were divided into training and internal validation sets at a ratio of 7:3 using stratified random sampling. The lesions from center 2 were included in the external validation set. Each lesion was regarded as an independent sample, and the sampling was performed to avoid any overlap of lesions between the two sets.

MRI acquisition

MRI examinations were performed using two 3.0 T MRI scanners (Magnetom Skyra, Siemens Healthineers, Erlangen, Germany, with an 18-channel body array coil; SIGNA Pioneer, GE Healthcare, Chicago, IL, USA, with a 32-channel body array coil). The imaging protocols used for transverse T2WI and DWI sequences are described in Table S1. ADC images were generated from DWI images using a monoexponential model: Sb/S0 = exp(−b × ADC), where S0 is the signal strength without diffusion sensitivity coefficient, Sb is the signal strength with diffusion sensitivity coefficient, and b is the diffusion sensitivity coefficient.

Histopathology analysis

Postoperative whole-mount sections of the prostate were prepared according to the methods of Liu (15). Histopathological sections were evaluated by a senior urogenital system pathologist with 13 years of experience. All clinically significant prostate cancer (csPCa) lesions were marked on the sections: GS ≥3+4, and/or pathological volume ≥0.5 mL, and/or capsule invasion. The location of the lesion in the PZ or TZ was judged based on the central position of the long axis.

Each tumor lesion was identified as sparse or dense based on its growth pattern through the semi-quantitative evaluation of the largest cross-sectional area (4,16,17): (I) lesion delineation: malignant areas separated by less than 3 mm apart in the same plane or adjacent slices were considered components of the same lesion (18,19); (II) component analysis: for the PCa lesion, each individual component was assessed. A sparse component was defined as a tumor region containing more than 60% of normal gland and stroma tissues. The remaining tumor tissue was classified as the dense component, characterized by a high proportion of malignant gland; (III) overall lesion classification: the entire PCa lesion was defined as “sparse” when the sparse components exceeded 50%; otherwise, the lesion was classified as “dense”.

Tumor segmentation

The three-dimensional tumor region of interest (ROI) was segmented slice by slice by a radiologist (reader 1) with 10 years of experience in prostate MRI on T2WI, DWI, and ADC images via ITK-SNAP software (version 3.8.0, http://www.itksnap.org) by referring to the anatomical characteristics according to the pathological sections (Figure 1), including PZ and TZ morphological features, location of the urethra and ejaculatory duct, and distance from the lesion to the top or base of the gland. To further assess the intra- and inter-observer repeatability, 25 patients randomly selected from the training cohort were reviewed, and the tumors were delineated again by reader 1 and another radiologist with 16 years of experience (reader 2) after eight weeks.

Figure 1 A 61-year-old patient with PCa with a PSA level of 17 ng/mL. (A) T2WI image showing a remarkably low signal lesion with obscure margin in the right PZ (long arrow) and a slightly low signal in the left PZ (short arrow). (B,C) DWI image showing a significantly high signal and low ADC value in the right PZ (long arrows), while no obvious abnormal signal or ADC value is found in the left PZ (short arrows). (D-F) H&E-stained histopathological slices (original magnification, ×10). The pathological slice shows PCa in both the right (long arrow) and left (short arrow) PZ (D). The lesion in the right PZ is composed of a high proportion of malignant gland. (E). The lesion in the left PZ consists of predominantly normal PZ tissue and a sparse cancer gland (F). PCa, prostate cancer; PSA, prostate specific antigen; T2WI, T2-weighted imaging; DWI, diffusion-weighted imaging; PZ, peripheral zone; ADC, apparent diffusion coefficient; H&E, hematoxylin and eosin.

The two radiologists scored each lesion according to the MRI findings and Prostate Imaging Reporting and Data System (PI-RADS) version 2.1. When there were differences in the lesion scores, the two radiologists reached an agreement through discussion.

Radiomics feature extraction and repeatability analysis

To correct for varying slice thicknesses and in-plane resolutions, each image was isotropically resampled into 1 mm³ voxels via the B-spline curve interpolation algorithm. Besides, images were discretized using a fixed bin width of 25 Hounsfield units. A total of 2,553 radiomics features were initially extracted from the original and wavelet images of the T2WI and DWI sequences and ADC maps using the open-access Python package PyRadiomics (version 3.0.1; Python Software Foundation, Wilmington, DE, USA), including first-order, shape, and texture features. The texture features included gray level cooccurrence matrix (GLCM), gray level size zone matrix (GLSZM), gray level run length matrix (GLRLM), neighboring gray tone difference matrix (NGTDM), and gray level dependence matrix (GLDM) features. The remaining parameters were configured using the default settings provided by the PyRadiomics package. More details about the radiomics feature extraction methods and descriptions can be found at https://pyradiomics.readthedocs.io/. Radiomics features with intra- and interclass correlation coefficient (ICC) values lower than 0.8, indicating poor agreement, were eliminated.

Radiomics feature selection

First, the extracted radiomics features were normalized using the z score method to eliminate scaling differences. Then, one-way analysis of variance (ANOVA) and least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation were used to identify the optimal features. In addition, the Kruskal-Wallis test was employed to evaluate the differences between the selected features from the sparse lesion, dense lesion, and noncancerous tissue groups.

Classification model construction and evaluation

The random forest classifier was used to construct a machine learning model for distinguishing dense lesions, sparse lesions, and noncancerous tissues, due to its robustness against overfitting and excellent generalization capability. Additionally, the class_weight parameter was set to “balanced” to tackle the class imbalance issue. For the TZ and PZ, a radiomics model and average ADC model were separately constructed based on the selected optimal radiomics features and average ADC values of the corresponding regions.

The classification performance of the established models was evaluated with respect to discrimination and calibration. First, receiver operating characteristic (ROC) curves, the AUC with 95% confidence interval (CI) and other quantitative metrics, such as precision, recall, and F1-score, were calculated. The DeLong test was further used to compare the differences in the AUC values among different models. Second, calibration curve plotting and Brier score calculation were applied to evaluate the consistency between the predicted and actual probabilities. An overview of the workflow used in this study is illustrated in Figure 2.

Figure 2 An overview of the workflow of the study. ****, P≤0.0001. ADC, apparent diffusion coefficient; ANOVA, analysis of variance; AUC, area under the curve; CI, confidence interval; DWI, diffusion-weighted imaging; ICC, intra- and inter-class correlation coefficient; LASSO, least absolute shrinkage and selection operator; PZ, peripheral zone; ROC, receiver operating characteristic; T2WI, T2-weighted imaging; TZ, transitional zone.

Statistical analysis

Variables among the three groups were analyzed using the Kruskal-Wallis test. The statistical contents related to optimization of radiomics features, model establishment, and model evaluation were detailed in Radiomics feature selection and classification model construction and evaluation sections. Statistical analysis was carried out based on R software (version 3.5.2; R Foundation for Statistical Computing, Vienna, Austria). A two-tailed P value less than 0.05 indicated statistical significance. The other data processing was implemented with Python programming language (version 3.8).

Results

Patient characteristics

In total, 156 patients with 372 PCa lesions were included in our final cohort (Figure 3, Table S2). The mean age of the patients was 68.25±5.92 years (range, 50–85 years). The median serum PSA level was 15.35±10.52 (IQR, 4.22, 48.30) ng/mL. The median GS was 4+4=8 (4+3=7, 4+4=8).

Figure 3 Flowchart of patient population. Center 1: Fudan University Shanghai Cancer Center; center 2: Zhongshan Hospital, Fudan University. csPCa, clinically significant prostate cancer; MRI, magnetic resonance imaging; RP, radical prostatectomy.

There were significant differences in the PI-RADS scores between sparse and dense PCa in all lesions, PZ lesions, and TZ lesions groups (P<0.001) (Table 1). The accuracies in diagnosing all, dense, and sparse lesions with a cutoff of 3 in the PZ were 55.87%, 97.97%, and 20%, respectively, whereas those in the TZ were 69.18%, 96.87%, and 26.98%, respectively.

The average ADC value showed significant differences among the three tissues (P<0.001). The average ADC values (×10⁻³ mm²/s) of noncancer tissues, dense lesions, and sparse lesions in the PZ were 1.60±0.37, 1.04±0.23, and 1.30±0.41, respectively, whereas those in the TZ were 1.38±0.27, 1.00±0.29, and 1.18±0.39, respectively.

Radiomics feature selection

Radiomics features were selected through repeatability analysis, ANOVA, and LASSO. Ultimately, the PZ and TZ lesions had 7 and 6 optimal features, respectively, as listed in Table S3. The boxplots in Figure S1 with statistically significant differences illustrate that the chosen features have high discriminability in distinguishing among noncancer tissues, dense lesions, and sparse lesions. Figure S2 shows the classification importance ranking and the relationship between the values and the effects of the model output, indicating the optimal features that played vital roles in the development of the random forest classifier.

Classification model evaluation

Figures 4,5 present the ROC curves of the established radiomics models and average ADC value models for the PZ and TZ, respectively, in the training, internal validation, and external validation cohorts. In terms of overall diagnostic efficacy (Table 2), in the PZ and TZ, the radiomics model showed significant improvement in micro-AUC values over the average ADC model (P<0.05). In terms of the diagnostic efficacy for each class (Table 3), the radiomics model performed significantly better than the average ADC model for diagnosing sparse lesions in all datasets (P<0.05), and for diagnosing dense lesions in the training and external validation sets (in the PZ). The overall quantitative metrics and quantitative metrics of each class calculated based on the classification results of the models were presented in Tables 4,5, respectively, indicating that the radiomics model was superior to the average ADC model. In addition, the calibration curve of the radiomics model indicated that it has good calibration, with low Brier scores (Figure 6) in diagnosing noncancerous tissues, sparse lesions, and dense lesions in the PZ and TZ.

Figure 4 The ROC curves of the established radiomics model for the PZ in the training (A), internal validation (B), and external validation (C) cohorts and for the TZ in the training (D), internal validation (E), and external validation (F) cohorts. The AUCs of the radiomics models in detecting sparse lesions, dense lesions, and noncancerous tissues, as well as their micro-average and macro-average values, were all above 0.88. AUC, area under the curve; CI, confidence interval; PZ, peripheral zone; ROC, receiver operating characteristic; TZ, transitional zone.

Figure 5 The ROC curves of the average ADC value model for the PZ in the training (A), internal validation (B), and external validation (C) cohorts and for the TZ in the training (D), internal validation (E), and external validation (F) cohorts. The AUCs of the average ADC value model in diagnosing dense lesions and noncancer tissues were similar, significantly higher than that in diagnosing sparse lesions. ADC, apparent diffusion coefficient; AUC, area under the curve; CI, confidence interval; PZ, peripheral zone; ROC, receiver operating characteristic; TZ, transitional zone.

Figure 6 Calibration curves in the training (A), internal validation (B), and external validation (C) cohorts show that the radiomics model achieved good calibration. PZ, peripheral zone; TZ, transitional zone.

Discussion

In this study, we explored the feasibility and value of radiomics based on bpMRI in the detection of dense and sparse PCa lesions and found that the radiomics model can effectively improve the detection of PCa lesions, especially sparse lesions.

According to PI-RADS version 2.1 (20), a score of 3 indicates an intermediate probability of malignancy, and a score of 4 or 5 indicates a high or very high probability. In previous studies, the diagnostic sensitivity of PI-RADS 3 was higher than 90%, and the AUC was approximately 0.85 (21). Therefore, PI-RADS 3 was defined as positive in this study. There was a significant difference in PI-RADS scores between sparse and dense lesions (P<0.001). The overall accuracy of the PI-RADS score in diagnosing PCa was 61.56%, and the accuracies of diagnosing dense PCa and sparse PCa were 97.43% and 22.47%, respectively. The PI-RADS score range for dense lesions is 3 to 5, making it easy to detect dense lesions on routine bpMRI. However, the PI-RADS scores of sparse lesions were mostly 1 and 2, which may cause misdiagnosis.

In this study, the average ADC value showed significant differences among the three tissues in the PZ and TZ (P<0.001); however, sparse lesions were closer to noncancer tissues, which was consistent with previous studies. Laudicella et al. (11) found that the average ADC (b-value 1,000 s/mm²) value of dilated growth lesions was (0.777±0.109) ×10⁻³ mm²/s, whereas the value of infiltrative growth lesions was (1.079±0.262) ×10⁻³ mm²/s. Langer et al. (4) found that the average ADC (b-value 600 s/mm²) value of noncancer tissues and dense lesions in the PZ was 1.46×10⁻³ and 1.15×10⁻³ mm²/s, respectively, whereas that of sparse lesions was similar to that of the noncancer tissues at 1.40×10⁻³ mm²/s. In previous studies and ours, the differences in ADC values of sparse and dense lesions may be related to differences in T2 relaxation time and b values (22). In addition, the T2 signal was significantly decreased in dense lesions compared with noncancerous tissue but not in sparse lesions. The characteristics of the ADC value and T2 signal make it difficult to diagnose sparse lesions. van Houdt et al. (5) classified PCa lesions into dense or moderate types based on the density of tumor glands and found that the T2 and ADC values of moderate type lesions were higher than those of dense type lesions; 21.8% of dense lesions and 65% of moderate lesions were missed on MRI. In this study, the AUCs (external validation dataset) of the average ADC value model of the sparse lesions and dense lesions were 0.62 (95% CI: 0.36–0.67) and 0.73 (95% CI: 0.55–0.74) in the PZ, 0.52 (95% CI: 0.31–0.60) and 0.83 (95% CI: 0.66–0.93) in the TZ, respectively, indicating that the average ADC value can effectively discriminate dense lesions while limit the detection of sparse lesions.

In previous studies, the ADC was shown to effectively detect the microenvironment of tumor tissue and changes in the epithelium, stroma, luminal space, and cell density and was therefore considered the best single parameter component for prostate MRI evaluation at present (23). Quantitative measurement of the average ADC value can significantly reduce the misclassification of lesions detected by MRI (24). The application of radiomics can further improve the accuracy of MRI diagnosis. Zhang et al. (25) attempted to use ADC values and MRI-based radiomics to predict endometrial cancer recurrence, with an average predictive efficiency of 0.709 for ADC values and 0.82 for radiomics. Fan et al. (26) found that the AUC of the minimal ADC value in the differential diagnosis of benign and malignant testicular masses was 0.767, whereas that of the radiomics model based on the ADC map was 0.868. The ADC-related artificial intelligence model developed by Sun et al. (27) displayed an accuracy of 0.849 in detecting PCa at the patient level. These values were similar to the findings of the current study, in which the average ADC value had good discrimination performance for PCa as a whole but poorly diagnosed sparse lesions. The performance metrics of the average ADC value model demonstrated consistently higher precision, recall, and F1-scores for diagnosing dense lesions compared to sparse lesions across the training cohort, internal and external validation cohorts. In contrast, the radiomics model exhibited variable performance across these lesion types with no consistent advantage for dense lesions over sparse lesions. Notably, within the internal validation cohort, the radiomics model achieved superior F1-scores for sparse lesions over dense lesions in both PZ and TZ analyses, indicating that the positive prediction result is more reliable. However, the F1-score of sparse lesions decreased in the external validation cohort, so further evaluation through multi-center studies is essential to validate the generalization of predictive performance across both dense and sparse lesions. From the overall and grouping point of view, the diagnostic performance of the radiomics model was higher than that of the average ADC value model, especially for sparse lesions. This suggests that the radiomics model may provide a new method for the diagnosis of PI-RADS 1–2 score or high ADC value lesions.

Although formal guidelines for sample size calculation in multiclass machine learning remain lacking, we addressed this limitation by adopting feature reduction and regularization techniques to control model complexity and assess robustness. The consistent model performance across both internal and external validation cohorts provide evidence that the current sample size is sufficient and the model is generalizable. Most of the radiomics features retained for model construction in this study came from the DWI or ADC sequences. The most important feature in the importance ranking was extracted from ADC sequence images of the PZ and DW images of the TZ. This is consistent with previous studies, such as the one by Xu et al. (13), who ultimately selected six radiomics features (two from DWI and four from ADC) based on bpMRI to establish a radiomics model to distinguish benign and malignant prostate lesions. In the diagnosis and differentiation of prostate cancer, DWI and ADC sequences can provide more information than T2WI. Wavelet features refer to the first-order and texture features calculated from an image after wavelet decomposition (28). Wavelet decomposition can change the size and distribution of pixels in the original image, resulting in changes in the obtained features and possibly generating new features. Yu et al. (29) found a significant correlation between wavelet features and the response to neoadjuvant therapy in a study predicting preoperative axillary lymph node metastasis using MRI-based radiomics. Zhou et al. (30) similarly found that wavelet features were better than non-wavelet features in predicting the efficacy of neoadjuvant therapy for locally advanced breast cancer. In this study, six wavelet features were used to construct the models, indicating that wavelet features can provide additional information to assist in diagnosis.

According to studies on false-negative and false-positive results of PI-RADS, approximately 15.8% of patients with PI-RADS 1–2 have csPCa confirmed by biopsy (31). This study may explain this phenomenon, as these lesions may be sparse and difficult to identify on conventional MRI. At the same time, the model constructed in this study can significantly improve the diagnosis of these lesions and can be used as an auxiliary screening for patients with PI-RADS 1–2. In addition, the risk stratification of PCa, such as identification of high-risk lesions, will facilitate adjustment of treatment options (32,33). With the development of radiomics and its combination with pathomics, PCa diagnosis and risk prediction can be more efficient (34-36).

This study has several limitations. First, this was a retrospective study, which determines its inherent inadequacy. Second, the sample size of this study was relatively small, and there may have been selection bias. Moving forward, we will extend the research timeframe and seek more cooperation from other centers to verify the reliability of the radiomics model. The correlation between the radiomics model and oncological outcomes will be further analyzed. Third, although ROIs were determined through consultation between the pathologist and radiologists, deviation may still exist, and automatic registration systems will be further studied for possible integration into our workflow.

Conclusions

This study validated the feasibility of radiomics model based on bpMRI to detect sparse and dense PCa lesions, which can improve the diagnostic shortcomings of PI-RADS score and average ADC values and can effectively improve the detection of PCa, especially sparse lesions, and significantly reduce missed diagnoses.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the CLEAR reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2912/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2912/dss

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2912/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This retrospective study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the ethics review board of Fudan University Shanghai Cancer Center (No. 2005217-Exp2), and individual consent for this analysis was waived due to the retrospective nature.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Kuhl CK, Bruhn R, Krämer N, Nebelung S, Heidenreich A, Schrading S. Abbreviated Biparametric Prostate MR Imaging in Men with Elevated Prostate-specific Antigen. Radiology 2017;285:493-505. [Crossref] [PubMed]
2. Brembilla G, Giganti F, Sidhu H, Imbriaco M, Mallett S, Stabile A, Freeman A, Ahmed HU, Moore C, Emberton M, Punwani S. Diagnostic Accuracy of Abbreviated Bi-Parametric MRI (a-bpMRI) for Prostate Cancer Detection and Screening: A Multi-Reader Study. Diagnostics (Basel) 2022;12:231. [Crossref] [PubMed]
3. Schieda N, Nisha Y, Hadziomerovic AR, Prabhakar S, Flood TA, Breau RH, McGrath TA, Ramsay T, Morash C. Comparison of Positive Predictive Values of Biparametric MRI and Multiparametric MRI-directed Transrectal US-guided Targeted Prostate Biopsy. Radiology 2024;311:e231383. [Crossref] [PubMed]
4. Langer DL, van der Kwast TH, Evans AJ, Sun L, Yaffe MJ, Trachtenberg J, Haider MA. Intermixed normal tissue within prostate cancer: effect on MR imaging measurements of apparent diffusion coefficient and T2--sparse versus dense cancers. Radiology 2008;249:900-8. [Crossref] [PubMed]
5. van Houdt PJ, Ghobadi G, Schoots IG, Heijmink SWTPJ, de Jong J, van der Poel HG, Pos FJ, Rylander S, Bentzen L, Haustermans K, van der Heide UA. Histopathological Features of MRI-Invisible Regions of Prostate Cancer Lesions. J Magn Reson Imaging 2020;51:1235-46. [Crossref] [PubMed]
6. Li EV, Schaeffer EM, Ramesh Kumar SKS, Zhou R, Yang XJ, Mana-Ay M, Vescovo M, Ho A, Keeter MK, Carr J, Casalino D, Kocherginsky M, Patel HD, Ross AE, Savas H. Utility of (18)F-DCFPyL PET for local staging for high or very high risk prostate cancer for patients undergoing radical prostatectomy. Eur J Nucl Med Mol Imaging 2025;52:2335-42. [Crossref] [PubMed]
7. Ferraro DA, Rüschoff JH, Muehlematter UJ, Kranzbühler B, Müller J, Messerli M, Husmann L, Hermanns T, Eberli D, Rupp NJ, Burger IA. Immunohistochemical PSMA expression patterns of primary prostate cancer tissue are associated with the detection rate of biochemical recurrence with (68)Ga-PSMA-11-PET. Theranostics 2020;10:6082-94. [Crossref] [PubMed]
8. Perera M, Papa N, Roberts M, Williams M, Udovicich C, Vela I, Christidis D, Bolton D, Hofman MS, Lawrentschuk N, Murphy DG. Gallium-68 Prostate-specific Membrane Antigen Positron Emission Tomography in Advanced Prostate Cancer-Updated Diagnostic Utility, Sensitivity, Specificity, and Distribution of Prostate-specific Membrane Antigen-avid Lesions: A Systematic Review and Meta-analysis. Eur Urol 2020;77:403-17. [Crossref] [PubMed]
9. Evangelista L, Zattoni F, Cassarino G, Artioli P, Cecchin D, Dal Moro F, Zucchetta P. PET/MRI in prostate cancer: a systematic review and meta-analysis. Eur J Nucl Med Mol Imaging 2021;48:859-73. [Crossref] [PubMed]
10. Rüschoff JH, Ferraro DA, Muehlematter UJ, Laudicella R, Hermanns T, Rodewald AK, Moch H, Eberli D, Burger IA, Rupp NJ. What’s behind 68Ga-PSMA-11 uptake in primary prostate cancer PET? Investigation of histopathological parameters and immunohistochemical PSMA expression patterns. Eur J Nucl Med Mol Imaging 2021;48:4042-53. [Crossref] [PubMed]
11. Laudicella R, Rüschoff JH, Ferraro DA, Brada MD, Hausmann D, Mebert I, Maurer A, Hermanns T, Eberli D, Rupp NJ, Burger IA. Infiltrative growth pattern of prostate cancer is associated with lower uptake on PSMA PET and reduced diffusion restriction on mpMRI. Eur J Nucl Med Mol Imaging 2022;49:3917-28. [Crossref] [PubMed]
12. Lambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J, Sanduleanu S, Larue RTHM, Even AJG, Jochems A, van Wijk Y, Woodruff H, van Soest J, Lustberg T, Roelofs E, van Elmpt W, Dekker A, Mottaghy FM, Wildberger JE, Walsh S. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14:749-62. [Crossref] [PubMed]
13. Xu M, Fang M, Zou J, Yang S, Yu D, Zhong L, Hu C, Zang Y, Dong D, Tian J, Fang X. Using biparametric MRI radiomics signature to differentiate between benign and malignant prostate lesions. Eur J Radiol 2019;114:38-44. [Crossref] [PubMed]
14. Prata F, Anceschi U, Cordelli E, Faiella E, Civitella A, Tuzzolo P, Iannuzzi A, Ragusa A, Esperto F, Prata SM, Sicilia R, Muto G, Grasso RF, Scarpa RM, Soda P, Simone G, Papalia R. Radiomic Machine-Learning Analysis of Multiparametric Magnetic Resonance Imaging in the Diagnosis of Clinically Significant Prostate Cancer: New Combination of Textural and Clinical Features. Curr Oncol 2023;30:2021-31. [Crossref] [PubMed]
15. Liu X, Zhou L, Peng W, Wang C, Wang H. Differentiation of central gland prostate cancer from benign prostatic hyperplasia using monoexponential and biexponential diffusion-weighted imaging. Magn Reson Imaging 2013;31:1318-24. [Crossref] [PubMed]
16. Starobinets O, Simko JP, Gibbons M, Kurhanewicz J, Carroll PR, Noworolski SM. The impact of benign tissue within cancerous regions in the prostate: Characterizing sparse and dense prostate cancers on whole-mount histopathology and on multiparametric MRI. Magn Reson Imaging 2024;114:110233. [Crossref] [PubMed]
17. Barral M, Jemal-Turki A, Beuvon F, Soyer P, Camparo P, Cornud F. Cellular density of low-grade transition zone prostate cancer: A limiting factor to correlate restricted diffusion with tumor aggressiveness. Eur J Radiol 2020;131:109230. [Crossref] [PubMed]
18. Ruijter ET, van de Kaa CA, Schalken JA, Debruyne FM, Ruiter DJ. Histological grade heterogeneity in multifocal prostate cancer. Biological and clinical implications. J Pathol 1996;180:295-9. [Crossref] [PubMed]
19. Villers A, McNeal JE, Freiha FS, Stamey TA. Multiple cancers in the prostate. Morphologic features of clinically recognized versus incidental tumors. Cancer 1992;70:2313-8. [Crossref] [PubMed]
20. American College of Radiology Website. Prostate Imaging Reporting & Data System (PI-RADS). 2019. Available online: https://www.acr.org/Clinical-Resources/Clinical-Tools-and-Reference/Reporting-and-Data-Systems/PI-RADS [accessed on 2020-12-2].
21. Tamada T, Kido A, Takeuchi M, Yamamoto A, Miyaji Y, Kanomata N, Sone T. Comparison of PI-RADS version 2 and PI-RADS version 2.1 for the detection of transition zone prostate cancer. Eur J Radiol 2019;121:108704. [Crossref] [PubMed]
22. Wáng YXJ. An explanation for the triphasic dependency of apparent diffusion coefficient (ADC) on T2 relaxation time: the multiple T2 compartments model. Quant Imaging Med Surg 2025;15:3779-91. [Crossref] [PubMed]
23. Chatterjee A, Watson G, Myint E, Sved P, McEntee M, Bourne R. Changes in Epithelium, Stroma, and Lumen Space Correlate More Strongly with Gleason Pattern and Are Stronger Predictors of Prostate ADC Changes than Cellularity Metrics. Radiology 2015;277:751-62. [Crossref] [PubMed]
24. Bonekamp D, Kohl S, Wiesenfarth M, Schelb P, Radtke JP, Götz M, Kickingereder P, Yaqubi K, Hitthaler B, Gählert N, Kuder TA, Deister F, Freitag M, Hohenfellner M, Hadaschik BA, Schlemmer HP, Maier-Hein KH. Radiomic Machine Learning for Characterization of Prostate Lesions with MRI: Comparison to ADC Values. Radiology 2018;289:128-37. [Crossref] [PubMed]
25. Zhang K, Zhang Y, Fang X, Dong J, Qian L. MRI-based radiomics and ADC values are related to recurrence of endometrial carcinoma: a preliminary analysis. BMC Cancer 2021;21:1266. [Crossref] [PubMed]
26. Fan C, Sun K, Min X, Cai W, Lv W, Ma X, Li Y, Chen C, Zhao P, Qiao J, Lu J, Guo Y, Xia L. Discriminating malignant from benign testicular masses using machine-learning based radiomics signature of appearance diffusion coefficient maps: Comparing with conventional mean and minimum ADC values. Eur J Radiol 2022;148:110158. [Crossref] [PubMed]
27. Sun Z, Wang K, Wu C, Chen Y, Kong Z, She L, Song B, Luo N, Wu P, Wang X, Zhang X, Wang X. Using an artificial intelligence model to detect and localize visible clinically significant prostate cancer in prostate magnetic resonance imaging: a multicenter external validation study. Quant Imaging Med Surg 2024;14:43-60. [Crossref] [PubMed]
28. Mallat SG. A Theory for Multiresolution Signal Decomposition:The Wavelet Representation. IEEE Trans Pattern Anal Mach Intell 1989;11:674-693.
29. Yu Y, He Z, Ouyang J, Tan Y, Chen Y, Gu Y, et al. Magnetic resonance imaging radiomics predicts preoperative axillary lymph node metastasis to support surgical decisions and is associated with tumor microenvironment in invasive breast cancer: A machine learning, multicenter study. EBioMedicine 2021;69:103460. [Crossref] [PubMed]
30. Zhou J, Lu J, Gao C, Zeng J, Zhou C, Lai X, Cai W, Xu M. Predicting the response to neoadjuvant chemotherapy for breast cancer: wavelet transforming radiomics in MRI. BMC Cancer 2020;20:100. [Crossref] [PubMed]
31. Salka B, Troost JP, Gaur S, Shankar PR, Diab AR, Hakim C, Mervak BM, Khalatbari S, Davenport MS. Clinical and Imaging Predictors of False-Positive and False-Negative Results in Prostate Multiparametric MRI Using PI-RADS Version 2. Radiol Imaging Cancer 2025;7:e240019. [Crossref] [PubMed]
32. Ferriero M, Prata F, Anceschi U, Astore S, Bove AM, Brassetti A, et al. Oncological Outcomes of Patients with High-Volume mCRPC: Results from a Longitudinal Real-Life Multicenter Cohort. Cancers (Basel) 2023;15:4809. [Crossref] [PubMed]
33. Ferriero M, Prata F, Mastroianni R, De Nunzio C, Tema G, Tuderti G, Bove AM, Anceschi U, Brassetti A, Misuraca L, Giacinti S, Calabrò F, Guaglianone S, Tubaro A, Papalia R, Leonardo C, Gallucci M, Simone G. The impact of locoregional treatments for metastatic castration resistant prostate cancer on disease progression: real life experience from a multicenter cohort. Prostate Cancer Prostatic Dis 2024;27:89-94. [Crossref] [PubMed]
34. Azadi Moghadam P, Bashashati A, Goldenberg SL. Artificial Intelligence and Pathomics: Prostate Cancer. Urol Clin North Am 2024;51:15-26. [Crossref] [PubMed]
35. Zhang YF, Zhou C, Guo S, Wang C, Yang J, Yang ZJ, Wang R, Zhang X, Zhou FH. Deep learning algorithm-based multimodal MRI radiomics and pathomics data improve prediction of bone metastases in primary prostate cancer. J Cancer Res Clin Oncol 2024;150:78. [Crossref] [PubMed]
36. Irmakci I, Nateghi R, Zhou R, Vescovo M, Saft M, Ross AE, Yang XJ, Cooper LA, Goldstein JA. Tissue contamination challenges the credibility of machine learning models in real world digital pathology. Modern Pathology 2024;37:100422. [Crossref] [PubMed]