Texture analysis of apparent diffusion coefficient maps for treatment response assessment in prostate cancer bone metastases—A pilot study

European Journal of Radiology 101 (2018) 184–190

Contents lists available at ScienceDirect

European Journal of Radiology journal homepage: www.elsevier.com/locate/ejrad

Research article

Texture analysis of apparent diffusion coefficient maps for treatment response assessment in prostate cancer bone metastases—A pilot study

T

⁎

Carolin Reischauera,b, , René Patzwahlc, Dow-Mu Kohd,e, Johannes M. Froehlicha, Andreas Gutzeita,f,g a

Institute of Radiology and Nuclear Medicine, Clinical Research Unit, Hirslanden Hospital St. Anna, Lucerne, Switzerland Institute for Biomedical Engineering, ETH and University of Zurich, Zurich, Switzerland c Department of Radiology, Cantonal Hospital Winterthur, Winterthur, Switzerland d Academic Department of Radiology, Royal Marsden NHS Foundation Trust, Sutton, Surrey, UK e CR-UK and EPSRC Cancer Imaging Centre, Institute of Cancer Research, Sutton, Surrey, UK f Department of Chemistry and Applied Biosciences, ETH Zurich, Zurich, Switzerland g Department of Radiology, Paracelsus Medical University Salzburg, Salzburg, Austria b

A R T I C L E I N F O

A B S T R A C T

Keywords: Prostate cancer bone metastases Apparent diffusion coefficient Treatment monitoring Treatment response Texture analysis

Objective: To evaluate whole-lesion volumetric texture analysis of apparent diffusion coefficient (ADC) maps for assessing treatment response in prostate cancer bone metastases. Materials and methods: Texture analysis is performed in 12 treatment-naïve patients with 34 metastases before treatment and at one, two, and three months after the initiation of androgen deprivation therapy. Four first-order and 19 second-order statistical texture features are computed on the ADC maps in each lesion at every time point. Repeatability, inter-patient variability, and changes in the feature values under therapy are investigated. Spearman rank’s correlation coefficients are calculated across time to demonstrate the relationship between the texture features and the serum prostate specific antigen (PSA) levels. Results: With few exceptions, the texture features exhibited moderate to high precision. At the same time, Friedman’s tests revealed that all first-order and second-order statistical texture features changed significantly in response to therapy. Thereby, the majority of texture features showed significant changes in their values at all post-treatment time points relative to baseline. Bivariate analysis detected significant correlations between the great majority of texture features and the serum PSA levels. Thereby, three first-order and six second-order statistical features showed strong correlations with the serum PSA levels across time. Conclusion: The findings in the present work indicate that whole-tumor volumetric texture analysis may be utilized for response assessment in prostate cancer bone metastases. The approach may be used as a complementary measure for treatment monitoring in conjunction with averaged ADC values.

1. Introduction Bone metastases occur in more than 90% of the cases of advanced prostate cancer [1]. Up to 80% of the patients initially respond well to androgen deprivation therapy (ADT) with a progression-free interval of 23–37 months [2] but thereafter develop resistance, which ultimately leads to tumor progression [3–5]. In clinical practice, response assessment typically relies on measurements of the prostate specific antigen (PSA) level. The parameter has not proven to be a surrogate biomarker for improved survival [6–8]. Standard imaging techniques such as technetium-99m bone scintigraphy and computed tomography which

are utilized to assess disease burden are incapable of evaluating biological activity over time [9]. For this reason, the response evaluation criteria of solid tumors regard bone metastases without associated softtissue masses as non-measurable [10]. In recent years, several studies have demonstrated the potential of diffusion-weighted imaging (DWI) for monitoring treatment response in bone metastases from advanced prostate cancer [11–15]. In responders to therapy, increased values of the mean and median apparent diffusion coefficients (ADCs) were thereby observed after the commencement of ADT relative to the baseline values. However, using this approach, the spatial heterogeneity of treatment response is neglected. To overcome

Abbreviations: ADC, apparent diffusion coefficient; DWI, diffusion, weighted imaging; GLCM, grey level co-occurrence matrix; MRI, magnetic resonance imaging; PSA, prostate specific antigen; ROI, region of interest ⁎ Corresponding author at: Hirslanden Hospital St. Anna, Institute of Radiology and Nuclear Medicine, Clinical Research Unit, St. Anna-Strasse 32, CH-6006 Lucerne, Switzerland. E-mail addresses: [email protected], [email protected] (C. Reischauer). https://doi.org/10.1016/j.ejrad.2018.02.024 Received 30 November 2017; Received in revised form 16 February 2018; Accepted 17 February 2018 0720-048X/ © 2018 Elsevier B.V. All rights reserved.

European Journal of Radiology 101 (2018) 184–190

C. Reischauer et al.

non-mono-exponential diffusion models for treatment monitoring in prostate cancer bone metastases [42].

this drawback, the use of functional diffusion maps [16] for response evaluation has been proposed [11,12,17] but the approach relies on an accurate co-registration of pre- and post-treatment ADC maps and the results may be sensitive to the choice of the threshold beyond which a significant ADC change is deemed to have occurred. Alternatively, treatment response of individual lesions in prostate cancer bone metastases may be evaluated using volumetric texture analysis on ADC maps. The approach permits evaluating signal heterogeneity of ADC estimates across lesions and may provide complementary information for response assessment. Texture analysis is a versatile technique that may be applied to any digital image advocating its use for medical image interpretation at an early stage [18–20]. Texture analysis provides information on the spatial variation of grey-level intensities or patterns, some of which are imperceptible to the human eye. Based on DWI data and the ADC, texture analysis has been widely employed for tumor characterization [21–24] and less frequently for prediction of treatment response [25–27]. According to the methods used to investigate the interrelationships of the voxels in the image, mathematical procedures to characterize texture fall into three main categories: syntactic, spectral, and statistical methods [28]. Depending on the number of voxels defining the local feature, the statistical approach can be further classified into first-order, second-order, and higher-order statistics. First-order statistical texture features provide information related to the grey-level distribution in the image but they neglect the relative positions of the various grey levels within the image. On the contrary, second-order statistics estimate properties of two voxel values occurring at specific locations relative to each other. Amongst other methods, in particular the grey level co-occurrence matrix (GLCM) proposed by Haralick et al. [29] has been widely utilized to characterize the spatial dependence of grey levels using second-order statistics. The aim of this exploratory study is to evaluate the potential of whole-lesion volumetric texture analysis of ADC maps for assessing treatment response in prostate cancer bone metastases, which may be used as a complementary measure for treatment monitoring in conjunction with averaged ADC values. To this end, texture analysis is performed in a treatment-naïve patient population with advanced prostate cancer before treatment begin and at one, two, and three months after the commencement of ADT. Repeatability, inter-patient variability, and changes under therapy of four first-order and 19 second-order statistical texture features derived from the GLCM are investigated. Beyond that, Spearman’s rank correlation coefficients are computed across time to delineate the relationship between the texture features and the serum PSA levels.

2.2. Diagnosis of bone metastases at skeletal scintigraphy Diagnosis of metastatic disease was made within seven days of the baseline MRI examination using whole-body skeletal radionuclide scintigraphy three hours after injection of 550 Mbq 99mTc-labeled dicarboxypropane diphosphonate (Tecleos, Cisbio, Gif-sur-Yvette, France). A dual-head whole-body scanner with a low energy high-spatial-resolution collimator (E-CAM, Siemens, Munich, Germany) was used for image acquisition. 2.3. Androgen deprivation therapy All patients underwent ADT by means of either medical treatment (three patients received goselerin (Zoladex®), two patients leuprorelin acetate (Lucrin®), two patients goselerin (Zoladex®) + denosumab (Xgeva®), two patients leuprorelin acetate (Lucrin®) + denosumab (Xgeva®), and one patient goselerin (Zoladex®) + bicalutamide (Casodex®) + denosumab (Xgeva®)) or surgical orchiectomy (two patients). ADT was commenced within two days of the baseline MRI examination. 2.4. MRI data acquisition MRI acquisition was performed a maximum of two days before and was repeated one, two, and three months after the initiation of ADT using a 1.5 T whole-body scanner (Achieva, release 3.2.3.4, Philips Healthcare, Best, the Netherlands). Axial images of the pelvis were acquired with the patient in the supine position using a four-element receive-only body coil array (Philips, Healthcare, Best, the Netherlands). DWI was performed using a single-shot spin-echo echoplanar imaging sequence with the following parameters: repetition time = 4506 ms, echo time = 63 ms, field of view = 400 × 256 mm2, voxel size = 2 × 2 mm2, 32 contiguous slices, slice thickness = 6 mm, signal averages = 5, parallel imaging reduction factor = 1.6, b = 0, 50, 100, 150, 200, 400, 600, 800 s/mm2. At baseline, the DWI scan was acquired twice to permit assessment of measurement repeatability. In addition, the imaging protocol included T1-weighted, T2-weighted, and proton density-weighted sequences. The total imaging time of each examination amounted to approximately 20 min. 2.5. Serum PSA measurements

2. Materials and methods

For treatment response assessment, serum PSA levels were determined in tandem with each MRI examination. A decline in the PSA level by more than 50% confirmed by a second measurement four weeks later was accepted as a response to treatment [30].

2.1. Patient population This prospective, single-center study was approved by the local Ethics Committee (Zurich, Switzerland) and informed written consent was obtained from all patients. Seventeen treatment-naïve patients who fulfilled all inclusion and exclusion criteria were recruited. The inclusion criteria were as follows: adults with histologically proven prostate cancer, evidence of bone metastases in the pelvis confirmed on skeletal scintigrams and no prior history of ADT, chemotherapy or radiation therapy. The exclusion were as follows: history of another malignancy, contraindication to magnetic resonance imaging (MRI) or unwillingness to participate in the clinical study. Three patients deceased prior to termination of the study and two patients withdrew willingness to participate in the study after an acute deterioration in their health condition and had to be secondarily excluded. As a result, twelve men (mean age = 76, range = 67–85) with a total of 34 pelvic bone metastases from prostate cancer (mean size = 7.64 cm3, range = 2.64–39.65 cm3) were included in the analysis. The patient population has previously been analyzed to assess the performance of

2.6. Diffusion data analysis and region of interest definition DWI data analysis was performed using in-house software written in MATLAB (The MathWorks, release 2017a, Natick, MA, USA). First, eddy current-induced image warping was corrected in the in-vivo datasets using a correlation-based affine registration algorithm. Second, perfusion-insensitive ADC maps were computed by fitting the monoexponential signal decay model to b-values > 200 s/mm2. Third, every post-treatment parameter map was co-registered to the corresponding pre-treatment parameter map using a robust multi-resolution alignment algorithm [31] that was extended to allow for affine transformations. Fourth, the ADC maps were linearly interpolated in the through-plane direction to yield isotropic voxel sizes prior to texture analysis. For each metastasis, regions of interest (ROIs) were manually drawn over each tumor-bearing slice on the pre-treatment ADC map by a single radiologist (AG) with 15 years of experience in DWI of the body, 185

European Journal of Radiology 101 (2018) 184–190

C. Reischauer et al.

2.7.2. Second-order texture analysis First, the original ADC values, were linearly transformed to grey levels g according to the following formula [32–34]:

g ⎧1 Ng − 1 ⎪ = 1 + ADCmax − ADCmin (ADC − ADCmin ) , ⎨ ⎪ Ng ⎩ (ADC < ADCmin )

(

)

(ADCmin ≤ ADC ≤ ADCmax ) ADC > ADCmax where Ng corresponds to the number of grey levels bins with g ∈ [1, Ng]. ADCmin and ADCmax denote the upper and lower limits of the ADC values. In the present work, these limits were either set to the minimum and maximum parameter values of each lesion (local paradigm) or the same limits were used for all metastases in the study (global paradigm) [33]. For the global paradigm, the limits were set such that the voxelwise ADC values of all lesions in the study fell within the defined range. To investigate the influence of the rebinning parameters on the results, texture analysis was performed using both the global and the local paradigm as well as different bin numbers Ng (16, 32, and 64 bins). Second, non-directional GLCMs were computed based on the manually segmented three-dimensional ROIs using a displacement vector with distance 1 and incorporating 26 neighboring voxel pairs in 13 independent directions [33]. The entry (i, j) in the GLCM thereby corresponds to the probability of occurrence of a voxel pair with grey levels i and j which are one voxel apart. Third, 19 second-order statistical texture features were derived from the non-directional GLCMs. These comprised the 14 texture features that were originally proposed by Haralick et al. [29] complemented by autocorrelation [35], dissimilarity [35], cluster shade [36], cluster prominence [36], and maximum probability [37].

Fig. 1. Illustrative diagram of the methodology for whole-tumor volumetric first-order and second-order statistical texture analysis based on ADC maps.

incorporating the diagnostic information of the skeletal scintigrams and the corresponding conventional anatomical MR images to simplify lesion localization. The two-dimensional ROIs were subsequently summed up, resulting in a three-dimensional ROI circumscribing the entire lesion. In so doing, a total of 34 ROIs (mean size = 955 voxels, range = 330–4956 voxels) were defined which were then applied to the ADC maps acquired at subsequent time points [12,17] since no changes in tumor volume were observed across the patient population.

2.8. Statistical analysis Statistical analysis of the data was performed in MATLAB (The MathWorks, release 2017a, Natick, MA, USA). Cohort baseline values for each texture feature were calculated and repeatability was quantified by the coefficient of variation defined as CV = 100%⋅ ( ∑ (σi/ Fi )2)/ N , where σi corresponds to the standard deviation of the feature in the ith lesion and Fi is its mean. N denotes the total number of metastases. Inter-patient variability was assessed by computing the variance of the mean of the two baseline estimates across all lesions. Changes in the feature values over time were evaluated using Friedman’s test, followed by pairwise comparisons of the feature values before and after the initiation of ADT. Using the Bonferroni correction, the significance level was set at p < 0.017 to adjust for multiple comparisons in the post-hoc tests. In all other statistical analyses, a p-value of p < 0.05 was considered statistically significant. No corrections for multiple comparisons were performed in these cases since the present study is mainly descriptive and many of the texture features likely exhibit a high degree of correlation. Furthermore, Spearman’s rank correlation coefficients were calculated across time to delineate the relationship between the various texture features and the serum PSA levels. To this end, weighted average values of each texture feature were computed on a per-patient basis in patients with multiple lesions. Correlations were considered strong if |r| ≥ 0.6, moderate if 0.4 < |ρ| < 0.6, and weak if |ρ| ≤ 0.4, respectively.

2.7. Texture analysis Volumetric texture analysis was performed using in-house software written in MATLAB (The MathWorks, release 2017a, Natick, MA, USA) based on previously published source code [32]. For each manually segmented three-dimensional ROI, first-order and second-order statistical texture features were computed based on the ADC maps before treatment begin and at one, two, and three months after the commencement of ADT. Fig. 1 schematically summarizes the methodology that was utilized. 2.7.1. First-order texture analysis For each metastasis at every time point, four first-order statistical texture features (skewness, kurtosis, energy, and entropy) were calculated from the ADC histograms. To investigate the influence of the number of histogram bins on the results, feature computation was performed thrice using 20, 40, and the optimum number of bins for each lesion at every time point. For the latter, the number of bins was computed according to the Freedman-Diaconis rule. 186

European Journal of Radiology 101 (2018) 184–190

C. Reischauer et al.

paradigm (ADCmin = 0.26·10−3 mm2/s and ADCmax = 1.8·10−3 mm2/s) and 32 bins were used for feature extraction. With the exception of cluster shade, the texture features showed moderate to high precision. Thereby, eight texture features exhibited coefficients of variation below 10% (lower part of Table 1). Friedman’s tests revealed significant changes in all feature values over time. The great majority of these changes were highly significant (p < 0.001) and in turn, 14 of 19 texture features exhibited significant changes in their values at all posttreatment time points relative to baseline (lower part of Table 2). Spearman’s rank correlation revealed weak to moderate correlations of most texture features with the serum PSA levels across time. Contrast, homogeneity, difference variance, difference entropy, maximal correlation coefficient, and dissimilarity exhibited strong correlations with the serum PSA levels (lower part of Table 3). When the local instead of the global paradigm was used for second-order statistical texture analysis, smaller percentage changes in the features values were observed under therapy on average. Nevertheless, Friedman’s tests detected significant changes in all texture features over time. Finally, when the local instead of the global paradigm was utilized, weaker correlations of the texture features with the serum PSA levels across time were observed on average. As for first-order statistical texture analysis, there were no substantial differences in the results when the number of bins Ng was varied.

Table 1 Average pre-treatment values of the ADC texture features as well as corresponding coefficients of variation and inter-patient variabilities. Feature

First-order statistics Skewness Kurtosis Energy Entropy Second-order statistics Energy Contrast Correlation Variance Homogeneity Sum average Sum variance Sum entropy Entropy Difference variance Difference entropy Inf. meas. of corr. I Inf. meas. of corr. II Maximal corr. coefficient Autocorrelation Dissimilarity Cluster shade Cluster prominence Maximum probability

Average baseline value

Coefficient of variation (%)

Inter-patient variability (%)

0.16 3.68 0.10 2.52

317.1 14.5 13.1 4.1

358.3 25.7 18.7 5.8

0.03 5.10 0.67 81.70 0.49 16.54 231.88 2.94 4.26 5.10 1.60 −0.16 0.72 0.09

24.6 16.8 8.8 12.4 6.5 6.8 14.3 4.4 4.8 16.8 4.6 13.4 5.8 12.3

55.7 52.5 16.4 52.4 16.8 23.9 60.1 11.9 13.6 52.5 13.9 25.4 11.3 26.3

79.17 1.57 118.02 6074.75 0.07

12.7 9.6 281.9 45.0 35.8

53.1 28.6 216.6 168.3 48.7

4. Discussion Various studies have shown the potential of DWI quantified by the ADC for treatment monitoring in prostate cancer bone metastases [11–15]. These studies have mostly relied on mean or median values averaged across entire lesions and have neglected the potential use of the signal heterogeneity of ADC estimates across lesions for treatment response assessment. The present preliminary study shows that this gap may be filled by the application of whole-lesion volumetric texture analysis, which may provide complementary information. First-order and second-order statistical texture features changed significantly in response to ADT. At the same time, the majority of texture features showed high precision and were correlated with the serum PSA levels across time. These findings suggest that whole-tumor volumetric texture analysis may provide unique information related to treatment response without the need for acquiring additional MRI sequences. The approach may be used as a complementary measure in conjunction with averaged ADC values for treatment monitoring in prostate cancer bone metastases under ADT but further studies are required to evaluate whether there is an incremental value of the technique compared to treatment monitoring based on lesion averaged ADC values only. The results demonstrate that whole-lesion volumetric texture analysis yields information related to signal heterogeneity of the ADC maps, which in turn reflects on the treatment response in prostate cancer bone metastases. Using first-order statistical texture analysis for instance, significantly decreased values of skewness were observed after the initiation of ADT relative to baseline. It can thus be concluded that the majority of voxels in the metastases featured increased ADC values under therapy while a minority showed unchanged or even decreased ADC values in response to treatment. Furthermore, using second-order statistical texture analysis, increasing values of contrast were observed at the same time. Contrast is a measure of the local ADC variations present in the tumor. Thus, the finding indicates that the voxels featuring unchanged or even decreasing ADC values under therapy were spread out throughout the lesions rather than being confined to one or several regions within the tumor. Nevertheless, it should be noted that many of the texture features are generally difficult to interpret. The present findings are an encouraging initial step but further studies are required to tap the potential of whole-tumor volumetric texture analysis for treatment monitoring in prostate cancer bone metastases. Volumetric texture analysis in conjunction with averaged ADC values may allow distinguishing between responders and non-

3. Results 3.1. Assessment of treatment response Changes in serum PSA levels under therapy confirmed treatment response in all patients. After initiation of ADT, average PSA values decreased by 85.7% (range = 51.4%–98.8%) at one month, 95.2% (range = 72.4%–99.9%) at two months, and 96.6% (range = 83.5%–99.9%) at three months after commencement of ADT. The serum PSA values of each individual patient can be found in the Supplemental Materials. In agreement with previous work [12–15], increasing mean tumor ADCs were simultaneously observed, in keeping with disease response. The mean ADC values of each lesion at every time point can be found in the Supplemental Materials. 3.2. First-order statistics The results of the first-order statistical texture analysis of the ADC maps are summarized in the upper parts of Tables 1–3. For calculation of the texture features, 20 bins were used for each metastasis at every time point. With the exception of skewness, all first-order statistical texture features exhibited high precision (upper part of Table 1). Friedman’s tests revealed significant changes in all four feature values over time. Thereby, the values of kurtosis, energy, and entropy were significantly altered at all post-treatment time points relative to baseline (upper part of Table 2). Bivariate analysis identified a strong correlation of the kurtosis, energy, and entropy values with the serum PSA levels across time (upper part of Table 3). There were no substantial differences in the results when first-order texture analysis was performed using the optimum number of bins according to the FreedmanDiaconis rule (mean size = 21, range = 8 − 59). 3.3. Second-order statistics The results of the second-order statistical texture analysis of the ADC maps are summarized in the lower parts of Tables 1–3. The global 187

European Journal of Radiology 101 (2018) 184–190

C. Reischauer et al.

Table 2 Results of the repeated measures analyses and pairwise comparisons of the pre-treatment values of the ADC texture features with those at one, two, and three months after onset of antiandrogen therapy. The values in parentheses are the percentage changes in the various feature values at each time point relative to baseline. Significant statistics in the post-hoc comparisons have p < 0.017 which incorporates a correction for multiple comparisons. Feature

Friedman test (p-value)

Pairwise comparisons (p-values) At 1 month after treatment begin

At 2 months after treatment begin

At 3 months after treatment begin

0.003* (−316.7) 0.001* (−15.6) < 0.001* (−16.2) < 0.001* (5.2)

0.027 (−238.9) 0.004* (−16.2) < 0.001* (−15.9) < 0.001* (5.2)

0.022 (−16.2) < 0.001* (87.6) 0.004* (9.2) < 0.001* (242.6) 0.003* (−13.9) < 0.001* (84.1) < 0.001* (297.5) < 0.001* (11.8) 0.002* (12.2) < 0.001* (86.7) < 0.001* (13.7) 0.008* (13.3) < 0.001* (8.5) 0.002* (−23.4) < 0.001* (247.6) < 0.001* (35.6) < 0.001* (−279.5) < 0.001* (245.5) 0.041 (−9.6)

0.005* (−28.2) < 0.001* (102.3) 0.092 (3.9) < 0.001* (202.2) < 0.001* (-16.0) < 0.001* (67.8) < 0.001* (248.3) < 0.001* (9.5) < 0.001* (11.0) < 0.001* (102.3) < 0.001* (15.5) 0.831 (0.2) 0.092 (3.5) < 0.001* (−20.8) < 0.001* (205.4) < 0.001* (40.4) 0.011* (−158.8) < 0.001* (202.6) 0.064 (−13.3)

First-order statistics Skewness Kurtosis Energy Entropy

0.007* 0.001* < 0.001* < 0.001*

0.001* 0.016* 0.001* 0.001*

Second-order statistics Energy Contrast Correlation Variance Homogeneity Sum average Sum variance Sum entropy Entropy Difference variance Difference entropy Inf. meas. of corr. I Inf. meas. of corr. II Maximal corr. coefficient Autocorrelation Dissimilarity Cluster shade Cluster prominence Maximum probability

< 0.001* < 0.001* 0.002* < 0.001* < 0.001* < 0.001* < 0.001* < 0.001* 0.001* < 0.001* < 0.001* < 0.001* < 0.001* < 0.001* < 0.001* < 0.001* < 0.001* < 0.001* 0.003*

0.006* (−27.2) < 0.001* (54.0) 0.006* (8.8) < 0.001* (242.0) 0.015* (−8.6) < 0.001* (86.4) < 0.001* (296.4) < 0.001* (9.1) 0.004* (8.7) < 0.001* (54.0) 0.009* (8.0) 0.001* (17.3) < 0.001* (8.9) 0.016* (−15.1) < 0.001* (248.0) 0.004* (21.1) < 0.001* (−239.5) 0.002* (152.3) 0.008* (−26.7)

(−316.8) (−11.6) (−12.3) (4.1)

cross validation. To evaluate the impact of the number of mathematical bins used to compute the results, statistical texture analysis was performed using varying bin numbers. Generally speaking, reducing the number of bins improves the counting statistics at the expense of discriminatory power [33] while the optimum number of bins depends on the lesion size [38]. Using different bin numbers, no substantial differences in the results of either first-order or second-order statistical texture analysis were observed which indicates that the textural features are relatively robust across a reasonable range of bin numbers. When the global rather than the local paradigm was utilized for second-order statistical texture analysis, higher percentage changes in the feature values were observed in response to therapy. Beyond that, stronger correlations of the texture features with the serum PSA levels across time were detected on average. Thus, whole-lesion volumetric texture analysis based on the global rather than the local paradigm may be more sensitive to changes in response to therapy. In principle, the global paradigm is generally preferred unless the local paradigm results in intensity limits of little variation across lesions and time points [33]. In the present work, statistical texture analysis was utilized to assess treatment response. Due to its widespread use, second-order statistical texture features were extracted using non-directional GLCMs. Alternatively, grey level run length matrices permit extracting secondorder statistical texture features [39]. However, either method may lack the sensitivity to identify larger scale or more coarse changes in spatial frequency [40]. These could be assessed using spectral features, which may thus provide complementary information for evaluating treatment response in prostate cancer bone metastases. There are limitations to the present study. First, only a relatively low number of patients could be included since most patients in this disease setting have already received either ADT, chemotherapy or radiation therapy prior to any imaging. Further studies including larger patient cohorts are needed to validate the current findings and to determine the incremental clinical value of whole-lesion volumetric texture analysis for treatment monitoring in prostate cancer patients.

Table 3 Spearman’s rank correlation coefficients and corresponding p-values between the various ADC texture features and the serum PSA levels. Feature

Spearman’s ρ

p-value

First-order statistics Skewness Kurtosis Energy Entropy

0.15 0.67 0.68 −0.64

0.291 < 0.001* < 0.001* < 0.001*

Second-order statistics Energy Contrast Correlation Variance Homogeneity Sum average Sum variance Sum entropy Entropy Diff. variance Diff. entropy Inf. Meas. Corr. I Inf. Meas. Corr. II Max. Corr. Coeff. Autocorrelation Dissimilarity Clus. Shade Clus. Prominence Max. probability

0.41 −0.72 −0.06 −0.35 0.67 −0.31 −0.35 −0.54 −0.54 −0.72 −0.72 −0.28 −0.05 0.60 −0.35 −0.73 0.39 −0.52 0.24

0.004* < 0.001* 0.673 0.014* < 0.001* 0.033* 0.015* < 0.001* < 0.001* < 0.001* < 0.001* 0.058 0.724 < 0.001* 0.014* < 0.001* 0.005* < 0.001* 0.094

responders to ADT and may even permit predicting the onset of resistance but further studies are essential to determine whether there is an incremental value of the technique compared to treatment monitoring based on lesion averaged ADC values only. To this end, longterm follow-up studies have to be performed which permit classification and evaluation of the feature space. Supervised classification techniques such as support vector machines or neural networks may then be utilized and accuracy or success of the strategy may be evaluated by 188

European Journal of Radiology 101 (2018) 184–190

C. Reischauer et al.

Radiology at the Cantonal Hospital Winterthur for their support during data acquisition. Finally yet importantly, we are grateful for the continuing support of Philips Healthcare, in particular of Dr. Markus Scheidegger and Erika Bruellmann, during the research project.

Second, for prove of principle, feasibility of texture analysis was evaluated in a treatment-naïve patient population within the first three months of ADT. From a clinical point of view, it would be invaluable to assess feasibility beyond this point, in particular in response to second line therapies such as radiotherapy and chemotherapy. Third, accuracy and reproducibility of ADC quantification and in turn texture analysis is affected by various acquisition parameters such as the repetition time, the echo time, the number and choice of b-values, and whether breathhold or free breathing techniques are being utilized. For this reason, recommendations on the use of DWI as a cancer imaging biomarker have been proposed [41]. Fourth, in line with previous work [12,17], three-dimensional ROIs circumscribing entire metastases were manually defined by a single radiologist on pre-treatment ADC maps. This approach is labor-intensive and may be compromised by limited interand intra-observer reproducibility. However, it has been shown that whole-lesion volumetric texture analysis outperforms two-dimensional texture analysis based on either rectangular or irregular ROIs containing the lesion on a representative slice [33]. Further studies are required to determine reproducibility of the technique based on manually defined ROIs or alternatively, automatic segmentation techniques may be utilized [14]. In conclusion, the present work demonstrates the potential of whole-tumor volumetric texture analysis of ADC maps for treatment response assessment in prostate cancer bone metastases under ADT. The findings show that the great majority of first-order and second-order statistical texture features yield high precision and change in response to therapy. Furthermore, significant correlations between the textural features and the serum PSA levels are observed across time. The method may be utilized as a complementary measure in conjunction with averaged ADC values for treatment monitoring in prostate cancer bone metastases but further studies are needed to evaluate whether there is an incremental value of the technique compared to treatment monitoring based on lesion averaged ADC values only.

Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at https://doi.org/10.1016/j.ejrad.2018.02.024. References [1] B.I. Carlin, G.L. Andriole, The natural history, skeletal complications, and management of bone metastases in patients with prostate carcinoma, Cancer 88 (S12) (2000) 2989–2994. [2] P.F. Schellhammer, R. Sharifi, N.L. Block, M.S. Soloway, P.M. Venner, A.L. Patterson, M.F. Sarosdy, N.J. Vogelzang, Y. Chen, G.J.C.M. Kolvenbag, et al., A controlled trial of bicalutamide versus flutamide, each in combination with luteinizing hormone-releasing hormone analogue therapy, in patients with advanced prostate carcinoma: analysis of time to progression, Cancer 78 (10) (1996) 2164–2169. [3] K.J. Pienta, D. Bradley, Mechanisms underlying the development of androgen-independent prostate cancer, Clin. Cancer Res. 12 (6) (2006) 1665–1671. [4] W.P. Harris, E.A. Mostaghel, P.S. Nelson, B. Montgomery, Androgen deprivation therapy: progress in understanding mechanisms of resistance and optimizing androgen depletion, Nat. Clin. Pract. Urol. 6 (2) (2009) 76–85. [5] T. Karantanos, P.G. Corn, T.C. Thompson, Prostate cancer progression after androgen deprivation therapy: mechanisms of castrate resistance and novel therapeutic approaches, Oncogene 32 (49) (2013) 5501–5511. [6] H.I. Scher, S. Halabi, I. Tannock, M. Morris, C.N. Sternberg, M.A. Carducci, M.A. Eisenberger, C. Higano, G.J. Bubley, R. Dreicer, D. Petrylak, P. Kantoff, E. Basch, W.K. Kelly, W.D. Figg, E.J. Small, T.M. Beer, G. Wilding, A. Martin, M. Hussain, G. Prostate Cancer Clinical Trials Working, Design and end points of clinical trials for patients with progressive prostate cancer and castrate levels of testosterone: recommendations of the Prostate Cancer Clinical Trials Working Group, J. Clin. Oncol. 26 (7) (2008) 1148–1159. [7] D.R. Berthold, G.R. Pond, M. Roessner, R. de Wit, M. Eisenberger, A.I. Tannock, T.A.X. investigators, Treatment of hormone-refractory prostate cancer with docetaxel or mitoxantrone: relationships between prostate-specific antigen, pain, and quality of life response and survival in the TAX-327 study, Clin. Cancer Res. 14 (9) (2008) 2763–2767. [8] S. Halabi, A.J. Armstrong, O. Sartor, J. de Bono, E. Kaplan, C.Y. Lin, N.C. Solomon, E.J. Small, Prostate-specific antigen changes as surrogate for overall survival in men with metastatic castration-resistant prostate cancer treated with second-line chemotherapy, J. Clin. Oncol. 31 (31) (2013) 3944–3950. [9] I. Jambor, A. Kuisma, S. Ramadan, R. Huovinen, M. Sandell, S. Kajander, J. Kemppainen, E. Kauppila, J. Auren, H. Merisaari, J. Saunavaara, T. Noponen, H. Minn, H.J. Aronen, M. Seppanen, Prospective evaluation of planar bone scintigraphy, SPECT, SPECT/CT, 18F-NaF PET/CT and whole body 1. 5T MRI, including DWI, for the detection of bone metastases in high risk breast and prostate cancer patients: SKELETA clinical trial, Acta Oncol. 55 (1) (2016) 59–67. [10] E.A. Eisenhauer, P. Therasse, J. Bogaerts, L.H. Schwartz, D. Sargent, R. Ford, J. Dancey, S. Arbuck, S. Gwyther, M. Mooney, L. Rubinstein, L. Shankar, L. Dodd, R. Kaplan, D. Lacombe, J. Verweij, New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1), Eur. J. Cancer 45 (2) (2009) 228–247. [11] K.C. Lee, D.A. Bradley, M. Hussain, C.R. Meyer, T.L. Chenevert, J.A. Jacobson, T.D. Johnson, C.J. Galban, A. Rehemtulla, K.J. Pienta, B.D. Ross, A feasibility study evaluating the functional diffusion map as a predictive imaging biomarker for detection of treatment response in a patient with metastatic prostate cancer to the bone, Neoplasia 9 (12) (2007) 1003–1011. [12] C. Reischauer, J.M. Froehlich, D.-M. Koh, N. Graf, C. Padevit, H. John, C.A. Binkert, P. Boesiger, A. Gutzeit, Bone metastases from prostate cancer: assessing treatment response by using diffusion-weighted imaging and functional diffusion maps–initial observations, Radiology 257 (2) (2010) 523–531. [13] C. Messiou, D.J. Collins, S. Giles, J.S. de Bono, D. Bianchini, N.M. de Souza, Assessing response in bone metastases in prostate cancer with diffusion weighted MRI, Eur. Radiol. 21 (10) (2011) 2169–2177. [14] M.D. Blackledge, D.J. Collins, N. Tunariu, M.R. Orton, A.R. Padhani, M.O. Leach, D.M. Koh, Assessment of treatment response by total tumor volume and global apparent diffusion coefficient using diffusion-weighted MRI in patients with metastatic bone disease: a feasibility study, PLoS One 9 (4) (2014) e91779. [15] R. Perez-Lopez, J. Mateo, H. Mossop, M.D. Blackledge, D.J. Collins, M. Rata, V.A. Morgan, A. Macdonald, S. Sandhu, D. Lorente, P. Rescigno, Z. Zafeiriou, D. Bianchini, N. Porta, E. Hall, M.O. Leach, J.S. de Bono, D.M. Koh, N. Tunariu, Diffusion-weighted imaging as a treatment response biomarker for evaluating bone metastases in prostate cancer: a pilot study, Radiology 283 (1) (2017) 168–177. [16] B.A. Moffat, T.L. Chenevert, T.S. Lawrence, C.R. Meyer, T.D. Johnson, Q. Dong, C. Tsien, S. Mukherji, D.J. Quint, S.S. Gebarski, P.L. Robertson, L.R. Junck, A. Rehemtulla, B.D. Ross, Functional diffusion map: a noninvasive MRI biomarker for early stratification of clinical brain tumor response, Proc. Natl. Acad. Sci. U. S. A. 102 (15) (2005) 5524–5529.

Funding The authors state that this work has not received any funding. Conflict of interest The authors state that there are no conflicts of interest. However, it should be noted that Dr. Froehlich is a consultant for Guerbet contrast media company but its products or services are not related to the subject matter of the article. Ethical statement This study was approved by the local Ethics Committee (Zurich, Switzerland) and informed written consent was obtained from all patients. Authors contribution Each author has participated sufficiently to take public responsibility for its content: study design: CR, D-MK, JMF, AG; data collection: RP, AG; data analysis: CR, data interpretation: CR, D-MK, JMF, AG, manuscript: CR, RP, D-MK, JMF, AG. Acknowledgements We thank our partner urologists, especially Dr. Christian Padevit from the department of Urology at the Cantonal Hospital Winterthur and Dr. Lukas Matter from the joint Urological Practice in Winterthur, for their continuous contribution in form of clinical work-up and research support. Furthermore, we want to thank the radiographers Barbara Eichenberger and Sara Schwarz from the Department of 189

European Journal of Radiology 101 (2018) 184–190

C. Reischauer et al.

[29] R.M. Haralick, K. Shanmugam, I. Dinstein, Textural features for image classification, IEEE Trans. Syst. Man. Cyb. Smc. 3 (6) (1973) 610–621. [30] G.J. Bubley, M. Carducci, W. Dahut, N. Dawson, D. Daliani, M. Eisenberger, W.D. Figg, B. Freidlin, S. Halabi, G. Hudes, M. Hussain, R. Kaplan, C. Myers, W. Oh, D.P. Petrylak, E. Reed, B. Roth, O. Sartor, H. Scher, J. Simons, V. Sinibaldi, E.J. Small, M.R. Smith, D.L. Trump, G. Wilding, et al., Eligibility and response guidelines for phase II clinical trials in androgen-independent prostate cancer: recommendations from the Prostate-Specific Antigen Working Group, J. Clin. Oncol. 17 (11) (1999) 3461–3467. [31] O. Nestares, D.J. Heeger, Robust multiresolution alignment of MRI brain volumes, Magn. Reson. Med. 43 (5) (2000) 705–715. [32] M. Vallieres, C.R. Freeman, S.R. Skamene, I. El Naqa, A radiomics model from joint FDG-PET and MRI texture features for the prediction of lung metastases in softtissue sarcomas of the extremities, Phys. Med. Biol. 60 (14) (2015) 5471–5496. [33] W. Chen, M.L. Giger, H. Li, U. Bick, G.M. Newstead, Volumetric texture analysis of breast lesions on contrast-enhanced magnetic resonance images, Magn. Reson. Med. 58 (3) (2007) 562–571. [34] Z. Li, Y. Mao, H. Li, G. Yu, H. Wan, B. Li, Differentiating brain metastases from different pathological types of lung cancers using texture analysis of T1 postcontrast MR, Magn. Reson. Med. 76 (5) (2016) 1410–1419. [35] L.K. Soh, C. Tsatsoulis, Texture analysis of SAR sea ice imagery using gray level cooccurrence matrices, IEEE Trans. Geosci. Remote Sens. 37 (2) (1999) 780–795. [36] R.W. Conners, M.M. Trivedi, C.A. Harlow, Segmentation of a high-resolution urban scene using texture operators, Comput. Vision Graph. 25 (3) (1984) 273–310. [37] R.M. Haralick, Statistical and structural approaches to texture, Proc. IEEE 67 (5) (1979) 786–804. [38] P. Gibbs, L.W. Turnbull, Textural analysis of contrast-enhanced MR images of the breast, Magn. Reson. Med. 50 (1) (2003) 92–98. [39] M.M. Galloway, Texture analysis using gray level run lengths, Comput. Graph. Image Process. 4 (2) (1975) 172–179. [40] H. Zhu, B.G. Goodyear, M.L. Lauzon, R.A. Brown, G.S. Mayer, A.G. Law, L. Mansinha, J.R. Mitchell, A new local multiscale Fourier analysis for medical imaging, Med. Phys. 30 (6) (2003) 1134–1141. [41] A.R. Padhani, G. Liu, D.M. Koh, T.L. Chenevert, H.C. Thoeny, T. Takahara, A. DzikJurasz, B.D. Ross, M. Van Cauteren, D. Collins, D.A. Hammoud, G.J.S. Rustin, B. Taouli, P.L. Choyke, Diffusion-weighted magnetic resonance imaging as a cancer biomarker: consensus and recommendations, Neoplasia 11 (2) (2009) 102–125. [42] C. Reischauer, et al., Non-mono-exponential analysis of diffusion-weighted imaging for treatment monitoring in prostate cancer bone metastases, Sci. Rep. 7 (1) (2017) 5809.

[17] C. Reischauer, D.-M. Koh, J.M. Froehlich, R.e. Patzwahl, C.A. Binkert, A. Gutzeit, Pilot study on the detection of antiandrogen resistance using serial diffusionweighted imaging of bone metastases in prostate cancer, J. Magn. Reson. Imaging 43 (6) (2016) 1407–1416. [18] R.A. Lerski, E. Barnett, P. Morley, P.R. Mills, G. Watkinson, R.N. MacSween, Computer analysis of ultrasonic signals in diffuse liver disease, Ultrasound Med. Biol. 5 (4) (1979) 341–350. [19] Y.P. Chien, K.S. Fu, Recognition of X-ray picture patterns, IEEE Trans. Syst. Man. Cyb. Mc. 4 (2) (1974) 145–156. [20] E.L. Hall, R.P. Kruger, S.J. Dwyer, D.L. Hall, R.W. Mclaren, G.S. Lodwick, A survey of preprocessing and feature extraction techniques for radiographic images, IEEE Trans. Comput. 20 (1971) 1032–1044. [21] R. Rozenberg, R.E. Thornhill, T.A. Flood, S.W. Hakim, C. Lim, N. Schieda, Wholetumor quantitative apparent diffusion coefficient histogram and texture analysis to predict gleason score upgrading in intermediate-risk 3 + 4 = 7 prostate cancer, AJR Am. J. Roentgenol. 206 (4) (2016) 775–782. [22] A. Wibmer, H. Hricak, T. Gondo, K. Matsumoto, H. Veeraraghavan, D. Fehr, J.T. Zheng, D. Goldman, C. Moskowitz, S.W. Fine, V.E. Reuter, J. Eastham, E. Sala, H.A. Vargas, Haralick texture analysis of prostate MRI: utility for differentiating non-cancerous prostate from prostate cancer and differentiating prostate cancers with different Gleason scores, Eur. Radiol. 25 (10) (2015) 2840–2850. [23] A.S. Kierans, H. Rusinek, A. Lee, M.B. Shaikh, M. Triolo, W.C. Huang, H. Chandarana, Textural differences in apparent diffusion coefficient between lowand high-stage clear cell renal cell carcinoma, AJR Am. J. Roentgenol. 203 (6) (2014) W637–44. [24] Y.J. Ryu, S.H. Choi, S.J. Park, T.J. Yun, J.H. Kim, C.H. Sohn, Glioma: application of whole-tumor texture analysis of diffusion-weighted imaging for the evaluation of tumor heterogeneity, PLoS One 9 (9) (2014) e108335. [25] Y. Lin, H. Li, Z. Chen, P. Ni, Q. Zhong, H. Huang, K. Sandrasegaran, Correlation of histogram analysis of apparent diffusion coefficient with uterine cervical pathologic finding, AJR Am. J. Roentgenol. 204 (5) (2015) 1125–1131. [26] P. Brynolfsson, D. Nilsson, R. Henriksson, J. Hauksson, M. Karlsson, A. Garpebring, R. Birgander, J. Trygg, T. Nyholm, T. Asklund, ADC texture—an imaging biomarker for high-grade glioma? Med. Phys. 41 (10) (2014) 101903. [27] C.N. De Cecco, M. Ciolina, D. Caruso, M. Rengo, B. Ganeshan, F.G. Meinel, D. Musio, F. De Felice, V. Tombolini, A. Laghi, Performance of diffusion-weighted imaging, perfusion imaging, and texture analysis in predicting tumoral response to neoadjuvant chemoradiotherapy in rectal cancer patients studied with 3T MR: initial experience, Abdom. Radiol. 41 (9) (2016) 1728–1735. [28] A. Kassner, R.E. Thornhill, Texture analysis: a review of neurologic MR imaging applications, AJNR Am. J. Neuroradiol. 31 (5) (2010) 809–816.

190