Home | About | Journals | Submit | Contact Us | Français |

**|**HHS Author Manuscripts**|**PMC4713240

Formats

Article sections

Authors

Related links

AJNR Am J Neuroradiol. Author manuscript; available in PMC 2016 July 1.

Published in final edited form as:

Published online 2015 October 15. doi: 10.3174/ajnr.A4534

PMCID: PMC4713240

NIHMSID: NIHMS712838

Joonsang Lee,^{1} Rajan Jain,^{2} Kamal Khalil,^{3} Brent Griffith,^{3} Ryan Bosca,^{4} Ganesh Rao,^{5} and Arvind Rao^{1,}^{*}

The publisher's final edited version of this article is available free at AJNR Am J Neuroradiol

See other articles in PMC that cite the published article.

Texture analysis has been applied to medical images to assist in tumor tissue classification and characterization. In this study, we obtained textural features from parametric (rCBV) maps of dynamic susceptibility contrast-enhanced magnetic resonance imaging images in glioblastoma and assessed their relationship with patient survival.

MR perfusion data of 24 patients with glioblastoma from the Cancer Genome Atlas was analyzed in this study. One- and two-dimensional texture feature ratios as well as kinetic textural features based on rCBV values in the contrast-enhancing lesion and the non-enhancing lesion of the tumor were obtained. Receiver Operating Characteristic, Kaplan-Meier analysis, and multivariate Cox proportional hazards regression analyses were used to assess the relationship between texture feature ratios and overall survival.

Several feature ratios are capable of stratifying survival in a statistically significant manner. These feature ratios correspond to homogeneity (p=0.008, based on log-rank test), angular second moment(p=0.003), inverse difference moment(p=0.013) and entropy(p=0.008). Multivariate Cox proportional hazards regression analysis showed that homogeneity, angular second moment, inverse difference moment, and entropy from the contrast-enhancing lesion are significantly associated with overall survival. For the non-enhancing lesion, skewness and variance ratios of rCBV texture were associated with OS in a statistically significant manner. For the kinetic texture analysis, the Haralick correlation feature showed a p-value close to 0.05.

Our study reveals that texture feature ratios from contrast-enhancing and non-enhancing lesion and kinetic texture analysis obtained from perfusion parametric maps provide useful information for predicting the survival in the patients with glioblastoma.

Glioblastoma multiforme (GBM) is one of the most common and aggressive types of malignant brain tumors. The prognosis for patients with GBM remains very poor with median survival rate between 12 and 15 months.^{1,2} Several computer-based analyses including image texture analysis have been proposed to improve the diagnostic performance of imaging-derived measurements in cancer studies including GBM.^{3} Image texture analysis measures the local characteristic pattern of image intensity and has been applied to different image processing domains, such as texture classification and texture segmentation to identify distinct textural regions in an image.^{4} In recent studies, texture analysis has been applied to medical images to assist in tumor tissue classification and characterization. One study of PET and CT showed that the features for tumor heterogeneity extracted from the normalized gray-level co-occurrence matrix (GLCM) could represent an independent prognostic predictor in patients.^{5} Another texture study in PET/CT suggested that regional and local characterization of the PET tracer heterogeneity in tumors is more powerful than global measurements currently used in clinical practice.^{6} Also, a textural feature study in non-small cell lung cancer showed that baseline ^{18}F-fluorodeoxyglucose (^{18}F-FDG) PET scan uptake values are associated with non-response to chemoradiotherapy.^{7} Recently, a novel method defined as textural kinetics was studied with breast DCE-MRI by S. Agner et al.^{8} This method attempted to capture spatiotemporal changes in breast lesion texture for classifying malignant and benign lesions.

In this work, we investigated tumor-derived texture feature ratios from rCBV values (derived from DSC-MRI) of two different tumor regions – the contrast-enhancing lesion (CEL) region and the non-enhancing lesion (NEL) region. We extracted first-order statistics such as homogeneity, mean, standard deviation, skewness, and kurtosis from the intensity histogram, as well as Haralick texture features obtained from the intensity GLCM.^{9} Subsequently, ratios of these texture features between Laplacian-of-Gaussian filtered and unfiltered versions of the rCBV map were also derived. Basically, the Laplacian-of-Gaussian filters are useful for detecting edges in images and the feature ratio can give us quantitative relation of features between filtered and unfiltered, which would provide an effective normalization to minimize the effects of any potential variations in MR images from various patients.^{10} Also, we obtained textural kinetic features of brain tumor dynamic susceptibility contrast (DSC) MRI data within these CEL/NEL regions-of-interest (ROIs). The purpose of this study was to determine the association of these DSC-MRI textural feature ratios with overall survival status of GBM.

We identified 24 TCGA GBM patients based on the availability of perfusion DSC-MRI data from The Cancer Imaging Archive (TCIA). One of patients has tumors in the left occipital region and left frontal region, respectively. These two tumors are treated distinctly. Previously, these data were assessed for genomic relationships with rCBV values.^{11} In this study, we performed one- and two-dimensional texture analysis as well as kinetic texture analysis of rCBV values within CEL and NEL regions for survival prediction. The clinical data were obtained from the cBioPortal for Cancer Genomics (http://www.cbioportal.org) (Table 1). In addition, a survival class variable was created by dichotomizing the overall survival value at 12 months based on the typical median survival time (12~15 months) in GBM.^{2,12}

Relative cerebral blood volume (rCBV) values were calculated from ROIs within the CEL, the NEL, and the normal-appearing white matter (NAWM), respectively, based on rCBV maps obtained previously.^{11} The methods for this processing are explained in deeper detail in Jain et al.^{11} The rCBV intensities for the CEL and NEL were normalized with the mean value of the rCBV intensities for the unaffected NAWM region.^{13} The ROIs of the CEL, NEL, and NAWM were segmented by experts manually after co-registering rCBV parametric maps with T1 post-contrast and T2 FLAIR images respectively. The NEL ROIs were placed adjacent to CEL margin in the white matter within FLAIR signal abnormality region. Figure 1 shows an example of an rCBV map from the tumor in a female patient.

Textural feature ratios were computed from the normalized rCBV data in two steps. First, we applied a Laplacian-of-Gaussian (Eq. [1]), ^{2}*G*, (LoG) filter to a normalized rCBV ROIs to obtain filtered images

$${\nabla}^{2}G(x,y)=\frac{-1}{\pi {\sigma}^{4}}(1-\frac{{x}^{2}+{y}^{2}}{2{\sigma}^{2}}){e}^{-({x}^{2}+{y}^{2})/2{\sigma}^{2}}$$

[1]

where σ corresponds to the standard deviation of the LoG filter (here we use a medium level of coarseness, σ =1.8).^{10,14} The filter size chosen is 11×11, which was determined from the standard deviation value. The LoG filter derives edge-like features from the local intensity variations in images. Gray-level co-occurrence matrices (GLCMs) were derived from both unfiltered and filtered images. Next, one- and two-dimensional textural features were computed from the GLCMs of the unfiltered and filtered images.^{14} Finally, ratios of filtered texture descriptors to the unfiltered texture descriptors were calculated to yield texture feature ratios.

Image gray-level heterogeneity was quantified by using first-order statistics such as mean, standard deviation, skewness, and kurtosis of the pixel intensity distribution. Skewness and kurtosis are a measure of the asymmetry and a measure of the peakedness of the distribution, respectively. For the two-dimensional texture features, in order to quantify the spatial distribution of the pixel values (rCBV values) within the ROI, we derived the GLCMs from the unfiltered and filtered images. The GLCM measures the probability of the occurrence of a specific gray-level value pair as a function of distance and direction. We used eight gray levels, commonly used in these types of studies^{15} with 1 pixel offset to compute the GLCMs from the filtered and original images. We then computed 13 different second-order Haralick statistical measures from the GLCMs^{17}. The detailed equations for the second-order texture features are described in the Appendix.

For the kinetic texture analysis, the gadolinium concentration time series of the DSC perfusion data in both CEL and NEL regions were extracted using an open-source software package: Quantitative Utility for Assessing TreatmenT RespOnse (QUATTRO).^{16} Each ROI voxel from the dynamic perfusion dataset was normalized by the corresponding mean NAWM intensity, and all 18 features discussed above were calculated for each time point in the DSC series. Thus, we have 18 kinetic texture features for each perfusion dataset. Each time series texture feature was then fitted to a third order polynomial model (Eq. [2]) to yield four coefficients (*b _{0}, b_{1}, b_{2}, b_{3}*)

$$f(t)={b}_{0}+{b}_{1}t+{b}_{2}{t}^{2}+{b}_{3}{t}^{3}$$

[2]

This four-dimensional coefficient vector was then projected to one-dimension using metric multidimensional scaling.^{17}

A total of 18 texture feature ratios were obtained and compared between the overall survival groups (> or ≤12 months). The predictive accuracy of the CEL and NEL texture feature ratios for survival status was assessed using the receiver operating characteristic (ROC) curve.

Correlations between 18 texture feature ratios & 18 kinetic texture features with associated p-values were assessed via Spearman’s rank correlation, which is a non-parametric measure of statistical dependence between two variables. Statistical significance was defined as p-value < 0.05. The Kruskal-Wallis test, a non-parametric method for testing the equality of population medians among groups, was used to determine whether the median feature value differed significantly between survival groups. Texture feature ratios were assessed with Kaplan-Meier and ROC analysis to measure their associations with overall survival. Each feature ratio was dichotomized based on an optimum cutoff value derived from ROC analysis. Survival difference between the groups was assessed via a log-rank test. Multivariate Cox proportional hazards regression analysis was performed to assess the texture feature ratios as predictors independent of volume, age, and Karnofsky performance status (KPS) for overall survival.^{18} In this study, the MATLAB v8.0 (The MathWorks Inc., Natick, MA) and R (R Project for Statistical Computing, http://www.r-project.org/) were used for statistical analyses.

The texture features with and without Laplacian-of-Gaussian filtration were obtained, and the ratios between the Laplacian-of-Gaussian filtered and unfiltered features were calculated for the CEL and NEL regions, respectively. For the CEL, there were strong positive correlations between homogeneity and IDM (r = 0.99, p < 0.001), and there were strong negative correlations between ASM and entropy (r = −0.94, p < 0.001). For the NEL, there were strong positive correlations between the variance and sum average (r = 0.94, p < 0.001) and between the variance and sum variance (r = 0.95, p < 0.001) and between the sum average and sum variance (r = 0.99, p < 0.001). The summary of the Spearman rank correlations and p-values for the CEL and NEL are listed in Table 2. Information about the four most significant feature ratios such as homogeneity, ASM, IDM, and entropy for the CEL and skewness, variance, sum average, and sum variance for the NEL are listed in Table 3.

Spearman’s rank correlation and associated p-value from the following feature ratios for the CEL region: homogeneity, ASM, IDM, and entropy. Below, the rank correlation and p-values for skewness, variance, sum average, and sum variance within **...**

Range in texture feature ratios for (i) homogeneity, ASM, IDM, and entropy with and without Laplacian-of-Gaussian filtration for CEL and (ii) skewness, variance, sum average, and sum variance with & without Laplacian-of-Gaussian filtration for **...**

The areas under the ROC curve (AUCs) for each significant predictor of 12-month survival status (survival class) and corresponding p-values were assessed and are summarized in Table 4. The AUCs for the CEL-derived feature ratios were 0.83 for homogeneity, 0.76 for ASM, 0.81 for IDM, and 0.80 for entropy. The AUCs for the NEL-derived feature ratios were 0.80 for skewness, 0.72 for variance, 0.72 for sum average, and 0.71 for sum variance. There was also a significant difference between survival classes for homogeneity (p = 0.008), ASM (p = 0.036), IDM (p = 0.013), and entropy (p = 0.015) from the CEL. However, no significant difference was found for the NEL-derived texture feature ratios (Table 5).

Areas under ROC curves (for prediction of 12 month survival status) from the CEL and NEL texture feature ratios. Only features with statistically significant AUCs are shown.

Kaplan-Meier survival curves for groups induced by the ROC-optimized cutoffs for the CEL-derived homogeneity, ASM, IDM, and entropy feature ratios are significantly different (p < 0.05) (Figure 2). The optimal cutoff points were 1.118 (p = 0.008) for homogeneity, 0.971 (p = 0.003) for ASM, 1.085 (p = 0.013) for IDM, and 1.00 (p = 0.008) for entropy. The median survival (in months) for each of the groups induced by the cutoff is listed in Table 6 for the CEL. Multivariate Cox proportional hazards regression analysis (including clinical variables like volume, age, KPS) showed that CEL-derived homogeneity, ASM, IDM, and entropy feature ratios had p-values of 0.004, 0.012, 0.006, and 0.0006, respectively, indicating that these feature ratios were independent predictors of overall survival. For the NEL, only skewness and variance feature ratios had p-values less than 0.05 (Table 7). From the kinetic texture analysis, only the Haralick correlation feature showed a p-value close to 0.05. All other features were not statistically significant (p > 0.1). Figure 3 shows the ROC curve for the kinetic Haralick correlation feature, with an AUC of 0.849 and p-value of 0.003 (Table 8).

Kaplan-Meier survival curves from ROC-induced cutoffs for CEL-derived feature ratios: (a) homogeneity, (b) ASM, (c) IDM, and (d) entropy.

ROC curve for prediction of survival status based on correlation feature from kinetic texture analysis. The AUC value was 0.849 and the 2.5% and 97.5% confidence intervals for the Mann-Whitney statistic were 0.667 and 0.952.

Kaplan-Meier analysis based on ROC for the CEL texture feature ratios (only significant features are shown)

Multivariate Cox Proportional Hazards Regression Analysis (in a model that includes
volume, age, KPS) for the CEL and NEL-derived rCBV texture feature ratios

Several studies have shown that the hemodynamic parameter rCBV from DSC-MRI is an important prognostic imaging biomarker that provides useful prognostic information in patients with GBM.^{11,19} Boxerman et al have shown that rCBV measurement is significantly correlated with GBM grade and can be used to predict time to progression and clinical outcome.^{13} Jain et al. showed that increased maximum rCBV in CEL is associated with increased risk of death and high rCBV in NEL and wild-type EGFR mutation are associated with poor survival.^{20} In our study, we applied texture analysis to the normalized hemodynamic parameter rCBV values from the ROIs of the CEL and NEL in the rCBV map to investigate the association of the perfusion MR-derived image textural feature ratios with overall survival in GBM.

Laplacian of Gaussian filter is a pre-calculated filter obtained from combining the Gaussian and Laplacian filters and is useful for detecting edges in images.^{10} The texture feature ratios in our study represent the quantitative relationship of features between the Laplacian of Gaussian filtered images and the unfiltered images^{14}. In a preliminary study, these feature ratios demonstrated lower dependence with scanner type compared to the original features. The purpose for the use of feature ratio (aside from following previous literature, such as Ganeshan et. al.)^{21} is to minimize the effects of any potential systematic variations in MR images from various patients across different scanning or acquisition protocols. With these texture feature ratios, statistically significant differences were found for CEL-derived homogeneity, ASM, IDM, and entropy feature ratios between survival classes that dichotomized survival at 12 months. This implies that these feature ratios are associated with overall survival rates of GBM.

First-order statistics such as standard deviation, skewness, and kurtosis describe the probability distributions of the pixel intensities, and second-order statistics, such as Haralick features, describe the spatial relationship between pairs of pixels. Tumor-derived pixel-based heterogeneity can be measured using first- and second-order statistics. Many researchers have sought to determine whether such heterogeneity is associated with malignancy.^{22} Several studies of ^{18}F-FDG PET/CT have suggested that tumor heterogeneity might provide better prognostic information, tissue characterization, and tumor segmentation.^{23}

Our results from the Kruskal-Wallis test indicate that the texture feature ratios for homogeneity (p = 0.008), ASM (p = 0.036), IDM (p = 0.013), and entropy (p = 0.015) from the CEL had a strong correlation with the survival group, suggesting that these texture feature ratios are associated with overall survival and could provide additional prognostic information. Also, the results of kinetic texture analysis showed that the correlation feature from kinetic texture analysis had a high predictive (AUC) value (0.85). Conversely, the texture feature ratios for the NEL exhibited no significant correlation with overall survival.

There were several limitations to address in our study. Firstly, this was a retrospective study performed on a publicly available patient subset, consisting of data acquired on multiple MRI systems with varying protocols. A study that evaluates the robustness of these feature ratios for the survival prediction task, across scanning protocols, scanner resolutions and a larger sample size is essential to establishing their predictive value. A large sample size study will also enable the application of appropriate multiple testing corrections to identify reliably predictive features (there is no correction for multiple testing in the current study because of its exploratory nature). Also, variable treatment regimens with surgery, radiation, and chemotherapy may have a confounding effect on the survival rates of the patients. A separate dataset with uniformity of treatment regimen is the next step to validating the predictive value of these feature ratios. Further, incorporating molecular markers like IDH mutation status or molecular subtype can be useful to assess the additional predictive value of imaging-based measurements to existing molecular markers. The inclusion of other modalities or contrasts such as T1 post-contrast or FLAIR MR image is a great avenue for future work as well.

Previous studies have suggested that textural features can be used in several areas of image analysis, such as segmentation, classification, and prediction of tissue abnormality. In this study, we found that several feature ratios obtained from the rCBV map, in addition to kinetic textures, provided useful information for predicting the 12-month survival status from the CEL and NEL regions of patients with GBM.

The methods developed in this work are sufficiently general and might be applicable to other disease processes and sites where perfusion MRI is used for assessment of disease or treatment response. Our study presents the results of an exploratory study demonstrating the relationship of texture feature ratios (from one- and two-dimensional texture features as well as kinetic texture features) with survival in GBM patients. These findings suggest that texture feature ratios from perfusion MRI data are a promising method as a clinical prognostic tool.

The authors acknowledge the support of NCI P30 CA016672, a Career Development Award from the Brain Tumor SPORE (to A.R.) and start-up funding from MD Anderson Cancer Center for this research. We thank Ms. Markeda Wade for scientific editing.

**Grant support:** We gratefully acknowledge the NCI Cancer Center Support Grant NCI P30 CA016672, a Career Development Award from the Brain Tumor SPORE P50CA127001-07 (to A.R) and start-up funding from MD Anderson Cancer Center to support JL’s research.

- DSC-MRI
- Dynamic Susceptibility Contrast-enhanced MRI
- GBM
- Glioblastoma multiforme
- GLCM
- Gray Level Co-occurrence Matrix
- CEL
- Contrast Enhancing Lesion
- NEL
- Non-Enhancing Lesion
- ASM
- Angular Second Moment
- IDM
- Inverse Difference Moment
- NAWM
- Normal-Appearing White Matter

In this appendix, we give detailed equations for two-dimensional texture features such as 13 Haralick texture features and homogeneity feature in equations [A1~A14].

The angular second moment measures (ASM) the homogeneity of an image. A more homogeneous image has fewer gray levels with higher pixel element of the GLCM and sum of square values:

$${f}_{1}=\sum _{i=1}^{{N}_{g}}\sum _{j=1}^{{N}_{g}}p{(i,j)}^{2},$$

[A1]

where *N _{g}* is the number of gray levels present in an image, and

Contrast measures the luminance (differences in gray-level intensity values) present in an image:

$${f}_{2}=\sum _{k=0}^{{N}_{g}-1}{k}^{2}(\sum _{i=1}^{{N}_{g}}\sum _{j=1}^{{N}_{g}}p(i,j)),k=|i-j|$$

[A2]

Correlation measures the gray-level linear dependence of pixels at specified positions:

$${f}_{3}=\frac{1}{{\sigma}_{x}{\sigma}_{y}}\sum _{i=1}^{{N}_{g}}\sum _{j=1}^{{N}_{g}}(ij)p(i,j)-{\mu}_{x}{\mu}_{y}$$

[A3]

Variance differentially weighs the gray levels that significantly deviate from the mean value of *p*(*i, j*):

$${f}_{4}=\sum _{i=1}^{{N}_{g}}\sum _{j=1}^{{N}_{g}}{(i-\mu )}^{2}p(i,j).$$

[A4]

The local homogeneity or inverse difference moment (IDM) enhances local homogeneous regions by reducing the weight of inhomogeneous regions where *i* ≠ *j*:

$${f}_{5}=\sum _{i=1}^{{N}_{g}}\sum _{j=1}^{{N}_{g}}\frac{1}{1+{(i-j)}^{2}}p(i,j),$$

[A5]

the sum and difference histograms form the principal axes of the second-order probability density function. The sum average [A10] and variance [A11] quantify the mean and extent of the sum histogram, respectively. The sum entropy [A12] and difference entropy [A15] measure the homogeneity of the sum and difference histograms, respectively.

Sum average:

$${f}_{6}=\sum _{k=0}^{2{N}_{g}-2}k\xb7{p}_{x+y}(k),$$

[A6]

Sum variance:

$${f}_{7}=\sum _{k=0}^{2{N}_{g}-2}{(k-{f}_{8})}^{2}{p}_{x+y}(k).$$

[A7]

Sum entropy:

$${f}_{8}=-\sum _{k=0}^{2{N}_{g}-2}{p}_{x+y}(k)\text{log}({P}_{x+y}(k))$$

[A8]

Entropy quantifies the homogeneity of the image, suggesting that homogeneous regions have lower entropy values:

$${f}_{9}=-\sum _{i=1}^{{N}_{g}}\sum _{j=1}^{{N}_{g}}p(i,j)\text{log}(p(i,j)).$$

[A9]

Difference variance:

$${f}_{10}=\sum _{k=0}^{{N}_{g}-1}\left[{(k-\sum _{k=0}^{{N}_{g}-1}l\xb7{P}_{|x-y|}(k))}^{2}\right]{p}_{|x+y|}$$

[A10]

Difference entropy:

$${f}_{11}=-\sum _{k=0}^{{N}_{g}-1}{p}_{|x-y|}(k)\text{log}({P}_{|x-y|}(k))$$

[A11]

Information measure of correlation I and II:

$${f}_{12}=\frac{({f}_{9}+\sum _{i=1}^{{N}_{g}}\sum _{j=1}^{{N}_{g}}{P}_{(i,j)}\text{log}[{p}_{(i)}{p}_{(j)}\left]\right)}{\sum _{g=1}^{{N}_{g}}{p}_{(g)}\text{log}\left[{p}_{(g)}\right]}$$

[A12]

$${f}_{13}=\sqrt{1-\text{exp}[-2|-\sum _{i=1}^{{N}_{g}}\sum _{j=1}^{{N}_{g}}{p}_{(i)}{p}_{(j)}\text{log}\left[{p}_{(i)}{p}_{(j)}\right]-{f}_{9}\left|\right]}$$

[A13]

In addition to the Haralick texture features, we added a homogeneity feature that measures the closeness of the distribution of elements in the GLCM to the GLCM diagonal.

$${f}_{14}=\sum _{i=1}^{{N}_{g}}\sum _{j=1}^{{N}_{g}}\frac{1}{1+(i-j)}p(i,j).$$

[A14]

1. Johnson DR, O'Neill BP. Glioblastoma survival in the United States before and during the temozolomide era. J Neuro-Oncol. 2012;107(2):359–364. [PubMed]

2. Mazurowski MA, Desjardins A, Malof JM. Imaging descriptors improve the predictive power of survival models for glioblastoma patients. Neuro Oncol. 2013;15(10):1389–1394. [PMC free article] [PubMed]

3. Chen W, Giger ML, Bick U, Newstead GM. Automatic identification and classification of characteristic kinetic curves of breast lesions on DCE-MRI. Medical physics. 2006;33(8):2878–2887. [PubMed]

4. Liu XW, Wang DL. Image and texture segmentation using local spectral histograms. Ieee T Image Process. 2006;15(10):3066–3077. [PubMed]

5. Cheng NM, Fang YH, Chang JT, et al. Textural features of pretreatment 18F-FDG PET/CT images: prognostic significance in patients with advanced T-stage oropharyngeal squamous cell carcinoma. Journal of nuclear medicine : official publication, Society of Nuclear Medicine. 2013;54(10):1703–1709. [PubMed]

6. Tixier F, Le Rest CC, Hatt M, et al. Intratumor heterogeneity characterized by textural features on baseline 18F-FDG PET images predicts response to concomitant radiochemotherapy in esophageal cancer. Journal of nuclear medicine : official publication, Society of Nuclear Medicine. 2011;52(3):369–378. [PMC free article] [PubMed]

7. Cook GJ, Yip C, Siddique M, et al. Are pretreatment 18F-FDG PET tumor textural features in non-small cell lung cancer associated with response and survival after chemoradiotherapy? Journal of nuclear medicine : official publication, Society of Nuclear Medicine. 2013;54(1):19–26. [PubMed]

8. Agner SC, Soman S, Libfeld E, et al. Textural kinetics: a novel dynamic contrast-enhanced (DCE)-MRI feature for breast lesion classification. Journal of digital imaging. 2011;24(3):446–463. [PMC free article] [PubMed]

9. Haralick RM, Shanmugam K, Dinstein IH. Textural features for image classification. Systems, Man and Cybernetics, IEEE Transactions on. 1973;(6):610–621.

10. Ganeshan B, Miles KA, Young RCD, Chatwin CR. Texture analysis in non-contrast enhanced CT: Impact of malignancy on texture in apparently disease-free areas of the liver. Eur J Radiol. 2009;70(1):101–110. [PubMed]

11. Jain R, Poisson L, Narang J, et al. Genomic mapping and survival prediction in glioblastoma: molecular subclassification strengthened by hemodynamic imaging biomarkers. Radiology. 2013;267(1):212–220. [PubMed]

12. Affronti ML, Heery CR, Herndon JE, et al. Overall Survival of Newly Diagnosed Glioblastoma Patients Receiving Carmustine Wafers Followed by Radiation and Concurrent Temozolomide Plus Rotational Multiagent Chemotherapy. Cancer. 2009;115(15):3501–3511. [PubMed]

13. Boxerman JL, Schmainda KM, Weisskoff RM. Relative cerebral blood volume maps corrected for contrast agent extravasation significantly correlate with glioma tumor grade, whereas uncorrected maps do not. AJNR. American journal of neuroradiology. 2006;27(4):859–867. [PubMed]

14. Ng F, Ganeshan B, Kozarski R, Miles KA, Goh V. Assessment of primary colorectal cancer heterogeneity by using whole-tumor texture analysis: contrast-enhanced CT texture as a biomarker of 5-year survival. Radiology. 2013;266(1):177–184. [PubMed]

15. Renzetti F, Zortea L. Use of a gray level co-occurrence matrix to characterize duplex stainless steel phases microstructure. Fracture and Structural Integrity. 2011;(16):43–51.

16. Bosca R, Johnson V, Jackson E. SU-E-QI-18: QUATTRO: An Open-Source Software Package for Quantitative Imaging Applications in Assessing Treatment Response. Medical physics. 2014;41(6):380–381.

17. Borg I, Groenen PJ. Modern multidimensional scaling: Theory and applications. Springer: 2005.

18. Zinn PO, Sathyan P, Mahajan B, et al. A novel volume-age-KPS (VAK) glioblastoma classification identifies a prognostic cognate microRNA-gene signature. PloS one. 2012;7(8):e41522. [PMC free article] [PubMed]

19. Law M, Young RJ, Babb JS, et al. Gliomas: predicting time to progression or survival with cerebral blood volume measurements at dynamic susceptibility-weighted contrast-enhanced perfusion MR imaging. Radiology. 2008;247(2):490–498. [PMC free article] [PubMed]

20. Jain R, Poisson LM, Gutman D, et al. Outcome prediction in patients with glioblastoma by using imaging, clinical, and genomic biomarkers: focus on the nonenhancing component of the tumor. Radiology. 2014;272(2):484–493. [PMC free article] [PubMed]

21. Ganeshan B, Panayiotou E, Burnand K, Dizdarevic S, Miles K. Tumour heterogeneity in non-small cell lung carcinoma assessed by CT texture analysis: a potential marker of survival. European radiology. 2012;22(4):796–802. [PubMed]

22. Bedard PL, Hansen AR, Ratain MJ, Siu LL. Tumour heterogeneity in the clinic. Nature. 2013;501(7467):355–364. [PMC free article] [PubMed]

23. Yu H, Caldwell C, Mah K, et al. Automated radiation targeting in head-and-neck cancer using region-based texture analysis of PET and CT images. International journal of radiation oncology, biology, physics. 2009;75(2):618–625. [PubMed]

PubMed Central Canada is a service of the Canadian Institutes of Health Research (CIHR) working in partnership with the National Research Council's national science library in cooperation with the National Center for Biotechnology Information at the U.S. National Library of Medicine(NCBI/NLM). It includes content provided to the PubMed Central International archive by participating publishers. |