Tumor spatial heterogeneity in myxoid-containing soft tissue using texture analysis of diffusion-weighted MRI

The objective of this study was to examine the tumor spatial heterogeneity in myxoid-containing soft-tissue tumors (STTs) using texture analysis of diffusion-weighted imaging (DWI). A total of 40 patients with myxoid-containing STTs (23 benign and 17 malignant) were included in this study. The region of interest (ROI) was manually drawn on the apparent diffusion coefficient (ADC) map. For texture analysis, the global (mean, standard deviation, skewness, and kurtosis), regional (intensity variability and size-zone variability), and local features (energy, entropy, correlation, contrast, homogeneity, variance, and maximum probability) were extracted from the ADC map. Student’s t-test was used to test the difference between group means. Analysis of covariance (ANCOVA) was performed with adjustments for age, sex, and tumor volume. The receiver operating characteristic (ROC) analysis was performed to compare diagnostic performances. Malignant myxoid-containing STTs had significantly higher kurtosis (P = 0.040), energy (P = 0.034), correlation (P<0.001), and homogeneity (P = 0.003), but significantly lower contrast (P<0.001) and variance (P = 0.001) compared with benign myxoid-containing STTs. Contrast showed the highest area under the curve (AUC = 0.923, P<0.001), sensitivity (94.12%), and specificity (86.96%). Our results reveal the potential utility of texture analysis of ADC maps for differentiating benign and malignant myxoid-containing STTs.


Introduction
Myxoid-containing soft-tissue tumors (STTs) are an uncommon, heterogeneous group of mesenchymal neoplasms with variable histologic features. They are characterized by the presence of abundant extracellular myxoid matrix secreted by the tumor cells. Myxoid matrix, which has very high water content, results in distinctive imaging features on magnetic resonance imaging (MRI) with extremely low and high signal intensity on T1-and T2-weighted sequences, respectively [1]. When characterizing myxoid-containing STTs, there is considerable overlap in the signal characteristics between benign and malignant myxoid-containing a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 STTs on the conventional T1-and T2-weighted MRI, which necessitates the use of biopsy for diagnostic confirmation.
In addition to conventional MRI, diffusion-weighted imaging (DWI) has recently been used to characterize soft tissue lesions. DWI can detect differences in the Brownian motion of water protons depending on variations in the tissue microstructure. For image quantification, the apparent diffusion coefficient (ADC) value, which has been widely studied as an imaging biomarker for tumor cellularity, can be estimated from at least two DWIs with different b values. It has been reported that DWI has the potential to differentiate benign and malignant STTs, but many discrepancies exist in the literature [2][3][4][5][6][7][8]. Van Rijswijk et al. [3], for example, reported that the mean ADC values of the malignant STTs were significantly lower than those of the benign STTs from 23 patients. Other authors [7] reported that there were no significant difference between the malignant and the benign STTs in their study of 29 patients, because of the overlap in the ADC values between the two groups. In a large study of 88 subjects, Nagata et al. [8] reported that mean ADC values for myxoid-containing STTs were significantly higher than those of their non-myxoid-containing counterparts, but they could not find significant difference in the mean ADC values between benign and malignant myxoid-containing STTs. According to the previous studies, the mean ADC value has limited ability to discriminate between benign and malignant myxoid-containing STTs since it can be affected by the extracellular matrix containing large amount of water as well as by tissue cellularity. Thus, more sophisticated parameter other than mean ADC value is required for differentiating benign and malignant myxoid-containing STTs.
Texture analysis is an emerging oncological imaging tool that provides information on intralesional heterogeneity without the need for additional image acquisition [9,10]. Significant heterogeneity exists within tumor due to regional variation in cell proliferation, cell death, metabolic activity, and vascular structure. Over the past few decades, numerous published articles have shown the ability of texture analysis to extract diagnostically valuable information from medical images obtained using various imaging modalities [11][12][13][14]. Many studies have reported promising results of texture analysis regarding diagnosis, grading, classification, and survival prediction of tumors in various organs [13][14][15][16][17][18]. Recently, encouraging results have emerged from preliminary studies on texture analysis of conventional MRI used to differentiate benign and malignant STTs [19,20]. Regarding texture analysis of DWI, there are a few studies which revealed potential value of these parameters as clinically relevant imaging marker in tumors of brain, breast, and ovaries [12,21,22].
To our knowledge, the usefulness of texture analysis of DWI in myxoid-containing STTs has not been fully examined. Little is known about the spatial heterogeneity in the myxoidcontaining STTs. In addition, there are no published studies regarding texture analysis of DWI for differentiating benign and malignant myxoid-containing STTs. Thus, the purpose of this study was to examine the spatial heterogeneity in myxoid-containing STTs, and to investigate the utility of DWI texture analysis for differentiating benign and malignant myxoid-containing STTs.

Study subjects
This retrospective study was approved by the Institutional Review Board of the Samsung Medical Center (IRB File No. 2015-12-023), which waived the requirement of informed consent. From June 2011 to June 2014, patients who were pathologically diagnosed with myxoid-containing STTs and underwent musculoskeletal MRI including DWI before the procedure were included in this study. Forty consecutive subjects (18 women, 22 men) with 40 myxoid-containing STTs (23 benign and 17 malignant myxoid-containing STTs) were included in the analyses. Six tumors, which were all benign, were pathologically confirmed by image-guided biopsy and 34 tumors were confirmed by surgical excision. The 23 benign myxoid-containing STTs consisted of schwannomas (n = 18), intramuscular myxomas (n = 3), and neurofibromas (n = 2) while the 17 malignant myxoid-containing STTs included myxoid liposarcomas (n = 10), myxofibrosarcomas (n = 3), malignant peripheral nerve sheath tumors (n = 2), and extraskeletal myxoid chondrosarcomas (n = 2).

Magnetic resonance imaging protocol
All 40 patients were examined using a 3.0-T MRI system (Achieva TX; Philips Healthcare, Best, The Netherlands). Various radiofrequency (RF) coils and MRI parameters were used based on the anatomical location of the lesions (Table 1). A single-shot spin-echo echo-planar DWI sequence was performed in the axial plane. Sensitizing diffusion gradients were sequentially applied in the x, y, and z directions with b values of 0, 400, and 800 s/mm 2 .

Image data analyses
For image analysis, all DWI data was transferred to a personal workstation and analyzed using in-house software written using MATLAB v. 7.6 (Mathworks, Natick, MA). Fig 1 shows the flow chart for image processing using DWI texture analysis in myxoid-containing STTs. ADC maps were computed by exponential fitting of the local signal intensities using three b values (0, 400, and 800 s/mm 2 ). The region of interest (ROI) was manually drawn on the ADC map by a single radiologist (H.S.K) with 11 years of experience in musculoskeletal radiology who was blinded to radiologic reports, clinical information, and histopathologic results. ROIs were drawn on each section of the ADC maps in order to contain the entire tumor. The tumor volume was calculated by multiplying the number of voxels within the ROI by the unit volume of a voxel. Texture analysis was performed by a radiology physicist (J.H.K) with 7 years of experience in radiology physics. For global features, a histogram was derived from the distribution of the voxels' ADC values within the ROI. From the histogram, we computed the mean (1 st statistical momentum), standard deviation (2 nd statistical momentum), skewness (3 rd statistical momentum), and kurtosis (4 th statistical momentum). For regional and local texture analysis, the ADC values within the ROI were resampled to yield 64 discrete values ranging from 1 to 64 (discrete ADC map). For regional features, the grey level size zone matrix's (GLSZM's) (m, n) was defined by the number of homogenous regions, given the homogeneous tumor size (n) and the intensity (m). From GLSZM, the intensity variability and size-zone variability were computed [23]. For local features, the grey level co-occurrence matrix (GLCM) was created for all 13 directions with a 1-voxel distance [24]. To create rotation-invariant GLCM, we averaged each GLCM direction, and from the averaged GLCM, 7 local features (energy, entropy, correlation, contrast, homogeneity, variance, and maximum probability) were computed. In summary, we extracted 4 global, 2 regional, and 7 local features from the ROI ( Table 2).

Statistical analyses
All statistical analyses were performed using SAS statistical software (version 9.4, SAS Institute, Cary, NC, USA) and MedCalc software (version 12.7.2, MedCalc, Mariakerke, Belgium). Normality of data distribution was assessed using the Kolmogorov-Smirnov test. Sex differences between the groups were tested by a Chi-square test. A student's t-test was used to test the difference between group means. Analysis of covariance (ANCOVA) was performed with adjustments for age, sex, and tumor volume. For the texture parameters that showed significant difference between benign and malignant STTs in ANCOVA, a receiver operation characteristic (ROC) curve was constructed and the area under the curve (AUC) and the corresponding P-values were calculated. Optimal cutoff value was determined by maximizing the Youden index (J = sensitivity + specificity-1) and 95% confidence interval (CI) was calculated using binomial exact. Sensitivity and specificity were calculated for the thus determined optimal cutoff values. The discriminatory power was classified as follows, based on AUC: 0.9-1 = excellent, 0.8-0.9 = good, 0.7-0.8 = fair, 0.6-0.7 = poor, 0.5-0.6 = failure. In addition, pairwise comparisons of AUC for all variables were performed and Bonferroni's correction was used to adjust P-value for multiple comparisons. A P-value less than 0.05 was considered statistically significant.

Results
The statistical significance of age, sex, tumor volume, and texture analysis parameters of ADCs between benign and malignant myxoid-containing STTs are summarized in Table 3.

Population characteristics and tumor volume
Sex ratios and ages were not significantly different between the two groups (P = 0.289 and 0.068); however the mean tumor volume of malignant myxoid-containing STTs was significantly higher than that of benign myxoid-containing STTs (P = 0.009).

Global features
Statistical analysis of global features showed that malignant myxoid-containing STTs had significantly higher kurtosis after adjustment for age, sex, and tumor volume (P = 0.040) when compared with benign myxoid-containing STTs. The P-value for kurtosis was not statistically significant before the adjustment. Other global features derived from the grey-level histogram, including mean ADC, were not significantly different between the two groups.

Regional features
Regional features, which included both intensity variability and size-zone variability, were significantly higher in malignant myxoid-containing STTs before adjusting for age, sex, and tumor volume (P = 0.026 and 0.011). However, neither of the regional parameters were significantly different between the two groups after adjustment (P = 0.548 and 0.258).

ROC analyses
The diagnostic performance of texture parameters that showed significant difference between benign and malignant myxoid-containing STTs in ANCOVA are summarized in Table 4. Contrast showed excellent discriminatory power with the highest AUC (0.923, 0.794-0.983 of 95% CI; P<0.001), sensitivity, and specificity. The optimal cut-off value was 57.199, and using  Fig 4). AUC of homogeneity and variance were significantly higher than those of kurtosis and energy, respectively.

Discussion
Differentiation of benign and malignant myxoid-containing STTs based on imaging findings are often challenging for radiologists due to significant overlap of radiologic findings between the two on both conventional and functional MR sequences. Final diagnosis usually relies on pathologic evaluation of tissue specimen, but biopsy may not be routinely performed for all tumors in clinical practice. Thus there has been a need for novel imaging biomarker that can help differentiate benign and malignant myxoid-containing STTs. Tumor heterogeneity of myxoid-containing STT In the current study, texture analysis based on ADC map was performed to examine the spatial heterogeneity of malignant and benign myxoid-containing STTs. We found significant difference between benign and malignant myxoid-containing STTs with regard to kurtosis computed from the histogram, and the energy, correlation, homogeneity, contrast, and variance computed from GLCM. As shown in Figs 2 and 3, the spatial pattern of the benign myxoid-containing STT is visually more heterogeneous than that of the malignant myxoid-containing STT   on the ADC map. The significant difference of quantified parameters obtained through texture analysis in our result is reflection of this difference of spatial pattern on ADC map. An important finding of our study is that a number of ADC texture analysis parameters can reveal the difference between two tumor groups which were not demonstrated in comparison of mean ADC value.
Regarding local features, our results showed that malignant myxoid-containing STTs had significantly higher energy, correlation, and homogeneity, but significantly lower contrast and variance when compared to those of benign myxoid-containing STTs. Energy refers to a degree of non-uniformity, while correlation refers to a degree of linearity. Furthermore, homogeneity is defined as a degree of closeness in the distributed GLCM values. Contrast refers to a degree of difference in the neighboring values, and variance refers to how far each value is from the mean in GLCM. Our local texture analysis result revealed that malignant myxoidcontaining STTs to be non-uniform, linear, clustered, similar, and narrower in distribution in GLCM when compared to benign myxoid-containing STTs. This result can be interpreted as malignant myxoid-containing STTs having a more spatially homogenous ADC map. ROC analyses revealed contrast, correlation, homogeneity, and variance to show overall good diagnostic performance. Especially, contrast showed excellent discriminatory power with the highest AUC value, sensitivity, and specificity. These promising results of local feature warrants further study for clinical application of theses texture parameters.
Global feature also revealed spatial pattern on the ADC maps of benign and malignant myxoid-containing STTs that were similar to those demonstrated for local feature. Malignant myxoid-containing STTs had significantly higher kurtosis after adjusting for age, sex, and tumor volume when compared with benign myxoid-containing STTs. A higher kurtosis indicates a sharp peak in the distribution of pixel values, which generally indicates a greater homogeneity of pixel values. In accordance with previously reported studies [8,25], there was no significant difference in mean ADC value between benign and malignant myxoid-containing STTs.
Previous studies using texture analysis of DWI showed significantly higher entropy in malignant tumors compared with those in benign tumors, suggesting that there is higher spatial heterogeneity in malignant tumors [12,22]. However, in this study, we found that malignant myxoid-containing STTs to be more spatially homogenous on ADC maps (higher kurtosis, energy, correlation, homogeneity, and lower contrast, variance in malignant myxoidcontaining STTs). Our findings, which are contrary to those of previous studies, may be due to the underlying histologic characteristics of myxoid-containing tumors. Unlike non-myxoid tumors, the abundant extracellular matrix is a major textural component of myxoid-containing STTs [25,26]. On T1-weighted image, schwannomas are reported to demonstrate inhomogeneous signal and myxofibrosarcoma are reported to demonstrate homogeneous signal [20]. However, little has been reported on degree of signal homogeneity in myxoid-containing STTs on DWI. The majority (20/23, 87%) of benign myxoid-containing STTs in our data set were neurogenic tumors, which are reported to have a characteristic fibrous center with a myxoid periphery [27,28]. This histologic composition may have contributed to increased heterogeneity in diffusivity of benign myxoid-containing STTs. Further studies evaluating the difference in diffusivity between benign and malignant myxoid-containing STTs at cellular level, including the interplay with extracellular matrix, should be performed.
Regional features, such as intensity variability, and size-zone variability, were computed from GLSZM [16]. For regional texture analysis, the ADC values within the ROI were converted into 64 different values ranging from 1 to 64 (2 6 ). In this study, GLSZM's (m, n) was defined by the number of homogenous clusters, given the tumor size of the clusters (n, maximum: the number of voxels within ROI) to their clusters' discrete values (m, maximum: 64).
The size-zone variability was computed by emphasizing the GLSZM in terms of the tumor size (n), while the intensity variability was calculated by emphasizing the GLSZM in terms of the discrete values (m). Therefore, the row of GLSZM increases as the number of voxels within the ROI (tumor size) increases, leading to high size-zone variability for a large tumor ROI. Statistical analysis showed that size-zone variability was a good texture parameter to differentiate between malignant and benign myxoid-containing STTs; however, we could not find any statistical significance in size-zone variability after adjusting for age, sex, and tumor volume (P = 0.258).
Besides the intrinsic limits of a retrospective study, there were several limitations in our work. First, the study population was relatively small with a limited number of histologic subtypes. In addition, we did not compare the diagnostic accuracy of texture analysis parameters with that of conventional MRI parameters. Additionally, we did not perform histologic correlation to clarify the discrepancies in texture analysis parameters between the two groups. Lastly, image data analyses were done by a single reader.
In conclusion, we found that the spatial patterns on the ADC map in malignant myxoidcontaining STTs are more homogenous than those in the benign myxoid-containing STTs, which is contrary to the previous studies mainly using non-myxoid tumors in other organs. ROC analyses revealed overall good diagnostic performance of local features suggesting the potential use of DWI texture analysis to classify benign and malignant myxoid-containing STTs. Further research on the association between tumor texture and underlying biological factors is needed.
Supporting information S1