Utility of T2-weighted MRI texture analysis in assessment of peripheral zone prostate cancer aggressiveness: a single-arm, multicenter study

T2-weighted (T2W) MRI provides high spatial resolution and tissue-specific contrast, but it is predominantly used for qualitative evaluation of prostate anatomy and anomalies. This retrospective multicenter study evaluated the potential of T2W image-derived textural features for quantitative assessment of peripheral zone prostate cancer (PCa) aggressiveness. A standardized preoperative multiparametric MRI was performed on 87 PCa patients across 6 institutions. T2W intensity and apparent diffusion coefficient (ADC) histogram, and T2W textural features were computed from tumor volumes annotated based on whole-mount histology. Spearman correlations were used to evaluate association between textural features and PCa grade groups (i.e. 1–5). Feature utility in differentiating and classifying low-(grade group 1) vs. intermediate/high-(grade group ≥ 2) aggressive cancers was evaluated using Mann–Whitney U-tests, and a support vector machine classifier employing “hold-one-institution-out” cross-validation scheme, respectively. Textural features indicating image homogeneity and disorder/complexity correlated significantly (p < 0.05) with PCa grade groups. In the intermediate/high-aggressive cancers, textural homogeneity and disorder/complexity were significantly lower and higher, respectively, compared to the low-aggressive cancers. 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. Texture analysis of T2W images provides quantitative information or features that are associated with peripheral zone PCa aggressiveness and can augment their classification.

MRI examination. The image acquisition protocol and/or settings were standardized across all centers, hence the term 'single-arm' in the title. All imaging was performed on 3T MRI systems (Siemens Healthineers) using standard vendor-supplied body and spine phased array coils for signal detection, without an endorectal coil. A minimum of four weeks was allowed between the last biopsy and MRI to avoid hemorrhage artifacts. The acquisition consisted of localizer scans, T 2 W, DW, DCE, and spectroscopic imaging. In this study, we utilized only the transverse T 2 W and DW images, which were acquired with a turbo spin-echo sequence (repetition/ echo time (TR/TE): 4000/101 ms; field of view (FOV): 200 × 200 mm; matrix: 320 × 320; slice thickness: 3 mm; interslice gap: 0.6 mm), and a single-shot echo-planar sequence with four b-values: 0, 100, 400, and 800 s/mm 2 (TR/TE: 3300/60 ms; FOV: 260 × 211 mm; matrix: 160 × 130; slice thickness: 3.6 mm), respectively. The images were oriented along the longest axis of the prostate, perpendicular to the urethra to best match routine histologic sectioning of the prostate. Pre-imaging preparations were performed in accordance with local institutional guidelines.
Histopathologic examination and tumor delineation. Patients underwent radical prostatectomy within 12 weeks after the MRI examination. The prostatectomy specimens were prepared locally according to histopathology protocols at each institution, which included fixation, serial sectioning (perpendicular to the urethra to facilitate spatial matching to MRI) into ~ 3-4 mm axial slices, and hematoxylin and eosin staining of microsections. An experienced local uro-pathologist examined the stained slides, outlined cancer foci, described cancer location, and graded them in accordance with the Gleason scoring system 1,2 .
The annotated whole-mount histology sections were visually matched to the T 2 W images based on anatomical landmarks such as urethra, ejaculatory ducts, size/shape of the peripheral zone and apex/base proximity. Moreover, descriptions from the pathology report were used as guidance. Tumor volumes of interest (VOIs) were then manually delineated (by Gabriel A. Nketiah, 5 years' experience in prostate MRI; and guided by a radiologist Jurgen J. Fütterer with > 11 years' experience) based on their location in histology and shape/appearance on the T 2 W images. The VOIs were subsequently transformed to the corresponding DW images via intensity-based rigid registration (Elastix toolbox 26 ) using Mattes mutual information similarity metric (Fig. 1a). This was done by first Scientific Reports | (2021) 11:2085 | https://doi.org/10.1038/s41598-021-81272-x www.nature.com/scientificreports/ co-registering the T 2 W images to the b = 0 s/mm 2 images, and then applying the resulting transformation to the VOI masks. The co-registrations were visually verified, and manually corrected in case of mis-registration, for instance due to geometric distortion on the DW images. Each tumor was assigned a grade group (GG) according to the ISUP prostate cancer grading system 2 , and then dichotomized into low-(GG 1) and intermediate/ high-(GG ≥ 2) aggressive cancers.
Post-processing and feature extraction. Two types of features were computed from the image VOIs: traditional intensity histogram features (number of features, n f = 11) from the T 2 W images and ADC maps, and second and high-order statistical image textural features (n f = 29) based on GLCM 17 and GLRLM 18 from the T 2 W images. The "2D average" approach 27 was employed to compute textural features from the tumor VOIs (Fig. 1b). First, the intensities within each VOI were discretized into 32 grey levels via fixed bin number quantization. The GLCMs and GLRLMs were computed per slice at one-pixel distance (∂ = 1) in four symmetric directions, θ = 0°, 45°, 90° and 135°. Textural features were computed from each directional matrix, and the mean of each feature across the slices was obtained. Finally, the average of each feature across the four directions was calculated to eliminate potential differences in directionality. This 2D approach was preferred over 3D texture analysis due to the presence of interslice gaps in our data acquisition. ADC maps were calculated from the nonzero b-value DW image datasets (100, 400, and 800 s/mm 2 ) by fitting a monoexponential decay model to the image intensities as a function of b-value in each voxel. ADC histogram features were then computed for each tumor VOI. The b = 0 s/ mm 2 image was excluded from ADC map computation to eliminate possible perfusion effects. The ADC was included as a benchmark metric for aggressiveness classification, since it has been previously shown to correlate with prostate cancer aggressiveness 10 .

Figure 1.
Illustration of tumor volume-of-interest (VOI) delineation and grey level co-occurrence matric (GLCM) texture feature extraction. (a) Histopathology slides were visually matched to the T 2 -weighted (T 2 W) images using anatomical landmarks. Peripheral (red) and transition (yellow) zone tumors were manually outlined on the T 2 W images based on their location on histology and appearance on T 2 W images, and then transformed to corresponding apparent diffusion coefficient (ADC) maps via registration or manually if the registration was deemed suboptimal. (b) Textural feature extraction from T 2 W VOIs using 2D average approach. For illustration purposes, the number of grey level bins was limited to four in this schematic, instead of the 32 used in actual computations. www.nature.com/scientificreports/ Prior to feature extraction, the T 2 W images were corrected for intensity non-uniformity using the N4 bias field correction algorithm 28 , and subsequently normalized using the automated dual-reference tissue normalization approach 29 . Briefly, two aggregate feature channel object detectors were separately trained to detect fat and muscle tissue regions, from which reference intensity values (90th and 10th percentiles, respectively) were calculated, and then utilized to normalize the 3D image intensities to pseudo T2 values by linearly scaling the reference values to their corresponding T2 values at 3T from literature. Unlike T 2 W image intensities, the ADCs were not normalized because they are quantitative measurements in nature. Also, outlier voxels within each VOI, defined as intensities outside the range [µ − 3σ, µ + 3σ] 30 were excluded; where μ and σ denote the mean and standard deviation of the intensities within each VOI. All features were computed in accordance with the image biomarker standardization initiative 27 . Statistical analysis and classification modeling. Spearman correlation coefficients were calculated to investigate associations between the T 2 W image features and prostate cancer grade groups (i.e. 1-5). Differences in feature values between the two aggressiveness classes (i.e. low versus intermediate/high) were evaluated using two-tailed Mann-Whitney U-tests. p values were corrected for multiple testing using Benjamini and Hochberg's approach 31 at false discovery rate of 0.05, with values < 0.05 considered statistically significant.
To evaluate the utility of the features in classifying the two cancer aggressiveness classes, a linear support vector machine (SVM) classifier was trained and tested separately for each feature set (i.e. ADC histogram, T 2 W histogram and T 2 W textural features), and the following combinations: T 2 W histogram + textural features, ADC histogram + T 2 W histogram, and ADC histogram + T 2 W histogram + textural features. In this analysis, we were particularly interested in how well cancer aggressiveness at one institution could be predicted by a model trained on data from the other institutions, using histogram features with and without textural features augmentation. For this, the classifier was iteratively (i = 1:number of institutions, n i ) trained and tested, each time using data from n i − i institutions as training set, with the i th institution held out as an independent external test set (Fig. 2). The training employed stratified 10-fold cross-validation for hyperparameter tuning and feature selection. Hyperparameter (misclassification cost, C) tuning and feature selection during training were performed concurrently via grid search over seven logarithmically spaced values between − 1 and 1 inclusive, and using recursive feature elimination 32 , respectively. The hyperparameter and feature sets with the lowest mean misclassification error over all 10-folds were selected to build the model. The cross-validation partitioning of the data during training was done on patient level rather than on tumor level to ensure that multiple tumors from the same patient were all either in the training or in the validation subset. Predictions on the test set (i.e. data from the hold out institution) were however done on tumor-level.
Receiver operating characteristic (ROC) curves were computed for each test set, from which the area under the curve (AUC) and 95% confidence intervals (95CI), accuracy, sensitivity and specificity were calculated to evaluate the performances of the classifiers (i.e. feature sets) across the centers. The optimal threshold for calculating the accuracy, sensitivity and specificity was determined from the training based on the Youden index 33 , and then applied to the test set.
The added value of T 2 W textural features was evaluated using two approaches. First, at each institution, the AUCs before (i.e. ADC histogram, T 2 W histogram, and ADC + T 2 W histogram) and after augmentation with T 2 W textural features (i.e. T 2 W histogram + textural features, and ADC histogram + T 2 W histogram + textural features) were compared using Delong's nonparametric approach for comparing the areas under two or more correlated receiver operating characteristic curves 34 . Secondly, the differences in performances across the institutions before and after augmentation with T 2 W textural features were compared using paired student t-test.
Prior to the classification modelling, two-way ANOVA was performed to evaluate potential effects of data origin (i.e. institution) and cancer aggressiveness on the features. Features for which the interaction between institution and cancer aggressiveness or main effect of institution were significant were excluded from the model. Each feature was log transformed to meet normality assumption requirement of ANOVA. The SVM classification

Results
Out of the 128 eligible patients, 32 patients were excluded due to unavailable MRI (n = 5), MRI artifacts and/or distortion (n = 4), no pathology report/grading (n = 15), unsatisfactory matching between histopathology and MRI (n = 8). Data of 96 patients (mean age = 61.3 ± 6.1 years) for whom good quality MRI and post-surgical histopathology data were available were included in this study. In these patients, 127 tumor volumes (mean [range] = 469 [101-1397] voxels) were identified, of which 104 were in the peripheral zone and 23 were in the transition zone. Figure 3 shows the flowchart of patient inclusion and exclusion. Due to the limited number of transition zone tumors, only the peripheral zone tumors (in 87 patients) were analyzed, of which 30 were stratified as low-aggressive, and 74 as intermediate/high-aggressive cancers. The overview of the characteristics of patients and tumors is given in Table 1.
Feature association with prostate cancer grade group. The Spearman correlation between the image features and cancer grade groups was significant (p < 0.05) for eight T 2 W intensity histogram and nine textural features (Table 2). Differences in features between the two cancer aggressiveness classes were significant for seven intensity histogram and 16 textural features. Generally, the textural features reflected higher disorder/ complexity (i.e. high entropy or lower homogeneity) in intermediate/high-aggressive tumors, and vice versa in low-aggressive tumors (Fig. 4). As expected, the majority of these features (7 intensity and 9 texture features) were found to be common between the two statistical tests. The two-way ANOVA showed no statistically significant interaction between the effects of data origin (institution) and cancer aggressiveness on any feature. Similarly, the main effect of institution was not significant. Table 3 and Fig. 5 show comparisons of the performance of the feature sets in SVM classification of the two aggressiveness classes at the various institutions. The added value of T 2 W textural features varied generally between the sites. At the individual centers, the augmentation of ADC and T 2 W histogram features with T 2 W textural features resulted in improvement in AUCs at four centers, though not statistically significant (p > 0.05). When considering the overall performance across the centers, the differences in AUCs before and after augmentation with textural features www.nature.com/scientificreports/ were not significant. However, the difference in accuracy was significant (p = 0.0218) for ADC + T 2 W histogram versus ADC histogram + T 2 W histogram + T 2 W texture features (Fig. 5). In terms of feature importance within the classifier, a similar trend as in the feature association with cancer grade group was observed. Textural features relating to similarity (grey level non-uniformity), maximum probability, and textural complexity (information measure of correlation) were the most frequently selected, in addition to minimum and 10 th percentile from intensity histogram (Table 4).

Discussion
T 2 W MRI provides high spatial resolution and tissue-specific contrast compared to DW and DCE imaging, but it is predominantly limited to qualitative radiological evaluation of the prostate. In a preliminary study using single-center data 24 , we showed that quantitatively derived T 2 W image textural features have the potential to serve as non-invasive markers for assessing aggressiveness. In this work, we extended and confirmed these findings in a multicenter cohort. T 2 W image textural features, particularly those reflecting homogeneity/similarity (angular second moment, run length non-uniformity, grey level non-uniformity), disorder (entropy) and textural complexity (information measure of correlation) correlated significantly with prostate cancer aggressiveness; and differed significantly between low-and intermediate/high-aggressive prostate cancers as defined by histopathology. Compared to the classifier based on the commonly used histogram metrics from ADC and T 2 W images, the classifier utilizing histogram features augmented with T 2 W textural features performed better, an indication that quantitative texture analysis of anatomical images has the potential to reveal additional morphological and pathophysiological information for radiomics-based assessment of prostate cancer aggressiveness. The usefulness of entropy/complexity and homogeneity associated textural features in prostate cancer aggressiveness assessment and classification was shown in our previous study 24 , and has also been reported by others 23,36,37 . Histologically, aggressive prostate cancers are characterized by poor differentiation, glandular structure deformation, and loss of cellular integrity of the prostate gland. This disrupts the tissue cyto-architectural patterns, potentially leading to decreased homogeneity and high disorder. Correlations between textural features and prognostic factors and clinical outcome have also been reported 36,38 . If validated, these quantitatively derived T 2 W image features could be combined with other MRI parameters as evidence-based markers for prostate cancer. In the context of this study setup, the findings could particularly be useful in active surveillance situations to follow-up on low-risk cancer patients thereby limiting the need for repeated biopsies.
Although a number of promising studies have reported the utility of MRI texture analysis in prostate cancer [19][20][21][22][23][36][37][38] , very few are based on multicenter cohorts 22 or focused on aggressiveness prediction/classification 19,21,23,36 . Multicenter data sharing is important to fulfill the high data demand for training radiomics-based decision support systems. Furthermore, multicenter studies are necessary to ascertain the applicability and robustness of texture analysis and radiomics, and to facilitate their clinical transition across centers. Texture analysis, which considers spatial relationships between pixels rather than individual pixel intensities as in a histogram, could possibly contribute to overcome the inter-institution and scanner variability challenges associated with multicenter data. Compared to DW and DCE imaging, T 2 W imaging is generally regarded as the most stable sequence in terms of imperviousness to scanner variations and gradient artifacts, and tolerance in patients (i.e. contrast agent-free). Although these factors add to its importance, T 2 W imaging is not currently used for quantitative assessment of prostate cancer aggressiveness mainly due to the non-quantitative nature of its signal intensities.
We found the classification performances of the individual feature sets (ADC histogram, T 2 W histogram and T 2 W texture) across the sites to be complementary and in some cases comparable, while the best performance was achieved when the feature sets were combined. The latter observation confirms the preliminary findings from our single-center study 24 . The improved classification performance of T 2 W intensity histogram features compared to some reported studies 19,37 may be attributable to the post-processing (i.e. intensity non-uniformity correction and standardization) of the T 2 W images, which ensured that the intensities were comparable and had consistent quantitative interpretation across the patients and institutions. Our results are consistent with the findings of  37 and Chen et al. 19 who also reported improved characterization of prostate cancer aggressiveness when using combined textural features from ADC and T 2 W images, compared to only mean ADC values or the individual feature sets respectively. On the contrary, Bonekamp et al. 20 found  39 . Surprisingly, in some instances the classifier performance was reduced when textural features were added. Whereas this observation is worth further investigation, we suppose it could be due to potential inefficiencies in the feature selection process. The goal of this study was to investigate the added value of T 2 W image textural features in prostate cancer aggressiveness assessment with respect to traditional histogram features. Hence, we opted to limit our texture feature extraction to this image sequence even though textural features can also be extracted from ADC maps. Similarly, a wide variety of other radiomic feature types exist that can be extracted including shape, Gabor, wavelet, grey level size zone matrix, neighbouring grey tone difference matrix, etc., features 40,41 . Our choice of using only GLCM and GLRLM features is not based on preference, but, these features have been extensively studied in different tissues and image modalities, and are generally regarded as intuitive. Nevertheless, including additional features may further enhance the classification of prostate cancer aggressiveness.
Feature repeatability and reproducibility are important aspects of radiomics pipeline, especially in a multicenter setting as in this study. Ideally, one would expect a good radiomic feature to be reproducible or stable across centers under the same conditions. A number of studies have recently investigated the repeatability of radiomic features mostly via test-retest analysis 39,[42][43][44] . However, there is no consensus on which feature sets are most repeatable, mainly because the features are heavily influenced by pre-processing configurations during extraction. Most studies 43,44 show that shape features are more repeatable. In our analysis we found neither interaction between the effects of data origin and aggressiveness nor the main effect of data origin to be significant on any feature, an indication that features were reproducible or stable across the institutions. Despite this, the classification accuracies varied generally within mean relative deviation [7-18%] across the institutions, which may suggest cross-site inconsistency in feature performance. Even though the two statistics (ANOVA & classification) are not directly comparable, a plausible explanation for this latter observation could be due to the highly imbalanced distribution of the data (i.e. classification classes) across the institutions. Even if significant effects were found in the former, it could be due to other potential confounding factors such as inherent differences in the cohort tumor heterogeneity across centers or possible inter-observer differences in the histopathologic grading that was used as ground truth, rather than the features. An alternative and possibly a more robust method of Our study data and hence methods had some limitations. Despite being a multicenter study, our cohort size is relatively small and the distribution of patients or cancer aggressiveness classes across the centers is imbalanced (Table 1). This could affect our study results, especially the significance of T 2 W textural features in the classification performance. For instance, in a similar classification task performed by Fehr et al. 37 , the added value of textural features was found to be significant only after high imbalances in data was corrected for via sample augmentation. Clinically, prostate cancers are usually categorized into three aggressiveness classes (low, intermediate and high) or more 46 . Due to our study cohort size, however, tumor aggressiveness was stratified into low and intermediate/high. Although this stratification is inadequate for prostate cancer management, it serves as an important basis in the disease management pathway and could also be used as a benchmark for further clinical evaluation in larger multicenter studies. For instance, it is important to rule out low-risk cancers not needing active treatment, which otherwise can be overtreated with associated side effects. The inadequate number of transition zone tumors for analysis limits the applicability of our results to only the peripheral zone. Cancers originating from the transition zone have different radiomic features 22 , they tend to be elusive and are primarily assessed based on their appearance on T 2 W imaging 8 . Hence, the utility of quantitatively derived T 2 W image textural features for transition zone tumor characterization would be of great interest. Machine learning classifier performances are typically evaluated based on their classification accuracy. In this study for instance, Table 3. Performance of ADC, T 2 W histogram and T 2 W textural features in SVM classification of low-(grade group 1) vs. intermediate/high-aggressive (grade group ≥ 2) prostate cancers at the various institutions. An SVM classifier was trained and tested for each individual institution, where the test data were from the institution being evaluated, and the training data were obtained from the remaining 5 institutions. ADC apparent diffusion coefficient, T 2 W T 2 -weighted, CI confidence interval, SVM support vector machine, RD relative deviation. *Nan indicates there is no interval. www.nature.com/scientificreports/ the added of value T 2 W textural features can be better appreciated when looking at accuracy scores. Imbalances in datasets that are unrepresentative of the overall population can therefore lead to inflated assessment of the classifier accuracy. However, this was taken into account during the training and prediction (i.e. balanced accuracy) by adjusting the sample weights to be inversely proportional to the class frequencies. Feature repeatability is an important aspect of radiomics pipeline, hence, the lack of this tests in our study is another limitation. Our study also lacks comparison with clinical readings involving PI-RADS scores, which were not available for this study. However, previous studies 19, 47 have shown radiomics-based models from combined ADC and T 2 W to outperform PI-RADS scores in prostate cancer detection and aggressiveness assessment. Our study setup was based on biopsy proven cases. The performance of textural features in prostate cancer detection and characterization remains to be evaluated prospectively without knowledge of lesion presence and location, especially as textural properties could be affected by inflammation or prostatitis. Also, in future multicenter studies, it would be interesting to perform PI-RADS scoring and VOI segmentation locally at each institution similarly to the histopathology evaluation. Ultimately, further studies in a much larger multicenter cohort are worth investigating.

Conclusion
This multicenter study confirms that texture analysis of T 2 W images provides quantitative information for assessment of peripheral zone prostate cancer aggressiveness. T 2 W MRI-derived textural features correlated significantly with pathological findings (cancer grade group) from multiple institutions and were sensitive to underlying pathological differences between low-and intermediate/high-grade prostate cancers in the peripheral zone. Although, we found the added value of T 2 W textural features in the classification of these cancer aggressiveness classes to be moderate, our study suggests that T 2 W textural features may have the potential to improve prostate cancer classification in multicenter settings. With a wide array of proposed radiomics features, further studies in larger multicenter cohort would be needed to ascertain their added value and robustness.