Quantitative variations in texture analysis features dependent on MRI scanning parameters: A phantom model

Abstract Objectives To evaluate the influence of MRI scanning parameters on texture analysis features. Methods Publicly available data from the Reference Image Database to Evaluate Therapy Response (RIDER) project sponsored by The Cancer Imaging Archive included MRIs on a phantom comprised of 18 25‐mm doped, gel‐filled tubes, and 1 20‐mm tube containing 0.25 mM Gd‐DTPA (EuroSpinII Test Object5, Diagnostic Sonar, Ltd, West Lothian, Scotland). MRIs performed on a 1.5 T GE HD, 1.5 T Siemens Espree (VB13), or 3.0 T GE HD with TwinSpeed gradients with an eight‐channel head coil included T1WIs with multiple flip angles (flip‐angle = 2,5,10,15,20,25,30), TR/TE = 4.09–5.47/0.90–1.35 ms, NEX = 1 and DCE with 30° flip‐angle, TR/TE=4.09–5.47/0.90–1.35, and NEX = 1,4. DICOM data were imported into an in‐house developed texture analysis program which extracted 41‐texture features including histogram, gray‐level co‐occurrence matrix (GLCM), and gray‐level run‐length (GLRL). Two‐tailed t tests, corrected for multiple comparisons (Q values) were calculated to compare changes in texture features with variations in MRI scanning parameters (magnet strength, flip‐angle, number of excitations (NEX), scanner platform). Results Significant differences were seen in histogram features (mean, median, standard deviation, range) with variations in NEX (Q = 0.003–0.045) and scanner platform (Q < 0.0001), GLCM features (entropy, contrast, energy, and homogeneity) with NEX (Q = 0.001–0.018) and scanner platform (Q < 0.0001), GLRL features (long‐run emphasis, high gray‐level run emphasis, high gray‐level emphasis) with magnet strength (Q = 0.0003), NEX (Q = 0.003–0.022) and scanner platform (Q < 0.0001). Conclusion Significant differences were seen in many texture features with variations in MRI acquisition emphasizing the need for standardized MRI technique.

Texture analysis describes the patterns of pixel intensity variations within an image calculated by a series of mathematical algorithms. 1 Numerous texture analysis features have been described in the literature and defined in the work of Haralick et al. 1 The use of a texture analysis applied to imaging studies including CT and MRI have been previously performed for the evaluation of multiple nonneoplastic disorders including the evaluation for mesial temporal sclerosis on MRI, 2 evaluation of intervertebral disc disease on MRI, 3 evaluation of hepatic fibrosis on both CT and MRI, 4-8 evaluation of subchondral bone on MRI. 9 Prior oncologic studies have also employed texture analyses to evaluate specific tumor features including the assessment of HPV status of oropharyngeal squamous cell carcinomas, 8 prognosis of head and neck neoplasms, [10][11][12] classification of gastric and colorectal tumors on CT, [13][14][15] genomic mapping and predictive marker identification of gliomas on MRI, [16][17][18][19] the identification of potentially prognostic predictors in lung cancer, 20,21 evaluation of genitourinary neoplasms on both CT and MRI, [22][23][24][25] and for the radiomic classifications of breast carcinoma subtypes. [26][27][28] In an effort to study specific patterns of tumor biology correlating with different imaging appearances, multi-institutional centers have worked toward pooling resources to make publicly available cancer imaging databases, such as The Cancer Imaging Archive (TCIA) and The Cancer Genomic Atlas (TCGA), to help facilitate research efforts in the arena of tumor genotype-phenotype analyses. 16,26,28 Prior research studies have used a radiomics approach for investigating prostate cancer radiotherapy responses, 25 responsiveness of neoadjuvant chemotherapy in breast cancer, 28 and prognostic predictions of advanced nasopharyngeal carcinoma. 12 However, larger studies and systematic reviews on radiomics have noted methodological variations as a source of difficulty precluding an accurate and collective interpretation of data. 11,29,30 Based on our knowledge of how changes in the CT scanning parameters varies texture analysis features 30 , as well as preliminary studies investigating the sensitivity of texture features to variations in MRI technique, 29,[31][32][33] we could similarly deduce that changes in MRI scanning parameters such as differences in magnet strength and scanner platform could also influence texture analysis features. Thus, the purpose of this study was to evaluate and quantify changes in MRI sequence parameters may have on texture analysis features using a simple, nonanatomic phantom model.

| MATERIALS AND METHODS
This study employed the use of a phantom for all image acquisitions, precluding the requirement for IRB approval.

2.A | Phantom development and MR imaging techniques
The construction of the phantom, and scan data of serial MRI scans of this phantom are publicly available as part of the Reference Image Database to Evaluate Therapy Response (RIDER) at The Cancer Imaging Archive (TCIA). 34 The original DICOM datasets and scan data on the RIDER phantom are available for public use in an effort to generate an initial consensus on how to harmonize the data collection and analysis for quantitative imaging methods applied to the measurement of drug and/or radiation treatment response. 35 The nonanatomic phantom used in the RIDER database was comprised of 18

2.B | DICOM segmentation and texture analysis
Original DICOM data sets were downloaded and then imported into in-house developed MATLAB (MathWorks, Natick, MA) texture analysis software to calculate texture analysis features. The texture analysis software was developed by the co-author (BL) and the use of this texture analysis program has been previously reported in the literature. 7,8,30 Image segmentation of phantom was performed manually by an experienced radiologist (co-author HK), using the same geometric boundaries and a uniform contour volume for each dataset in an effort to reduce potential variation related to the manual segmentation process. The entirety of the phantom was contoured including each of the doped gel-filled tubes, the gadolinium filled tube, as well as the negative space between in the inserts. A correction for spatial inhomogeneity was not applied. Prior to the texture analysis, the contoured images were preprocessed (or corrected) which consisted of the following steps: (a) partial volume artifact correction, and (b) global grayscale normalization. These steps are described in the work by Li et al. 7 In brief, to correct for partial volume artifact, an optimal thresholding algorithm was applied using an iterative optimal thresholding algorithm. 36 This method assumes all image pixels are from two probability distributions (e.g., structure of interest and the dark background) and attempts to find the gray-level threshold corresponding to the minimum probability between the maxima of the two distributions, which results in minimal segmentation error. To find the optimal threshold, this algorithm was applied iteratively (usually four to ten iterations were sufficient), updating the threshold in each iteration from the weighted sum of the two distributions. For global grayscale normalization, the images were corrected by the mean and standard deviation to minimize the overall grayscale variation across images, similar to that described in the work of Collewet et al. 33 The correction was applied to the entire image. The mean gray value of each corrected image was set to 250 and the standard deviation to 30.
In total, 41 texture features, including 12 histogram features, five gray-level co-occurrence matrix (GLCM) features, 11 gray-level runlength (GLRL) features, four gray-level gradient matrix (GLGM) features, and nine Laws features, were calculated and averaged over the contoured images of each dataset. Numerous texture analysis equations have been defined and developed. Only a subset of 41 texture features were employed in this study based on our prior work, and based on the popularity of reported texture features in the radiomics literature. 7,8,30 The use of our in-house developed MATLAB program and the specific details of the texture analysis features calculated by this program have been previously published. 7,8,30 A full description of the mathematical equations is described in the work by Haralick et al. 1 and Tang el al. 37 GLCM features, in contrast to histogram features, are highly spatially dependent. In this study, the GLCM texture features were calculated using only directly adjacent pixels for simplicity. Horizontal, 45°, vertical, and 135°directions were averaged together to eliminate any directional dependence. The following GLCM features proposed by Haralick et al. 1 were tested: Angular Second Moment (ASM) ¼ ∑ i;j pði; jÞ 2 (3) where (i, j) represents the (i, j) value of the GLCM.
GLRL matrices were used as these texture features provide additional insights into spatial dependence18. The same directions consid- LGRE where p(i, j) represents the (i, j) value of the GLRL matrix, nr is the total number of runs, and np is the total number of pixels.
GLGM features were also investigated to provide the histogram of the absolute gradient values in the interrogated region of interest.
As a preprocessing step, the gradient of each pixel within the ROI was computed using a 3 × 3 neighborhood. The GLGM features mathematically summarize the gradient values of the pixels in the ROI and include mean, variance, skewness, and kurtosis.

| RESULTS
Changes in texture analysis features based on variations in MR scanning parameters are shown in Tables 1-4, and Table S1.

3.A | Assessment of magnetic strength
Variations in magnetic strength (1.5 T vs 3 T) resulting in changes in texture features are displayed in Table 1. No statistically significant differences were noticed in the histogram, or GLCM texture features. Mean texture analysis features on a 1.5 T vs a 3.0 T scanner. n: number of contoured slices; STD: standard deviation; STD5: 5-neighborhood standard deviation; STD9: 9-neighborhood standard deviation; IQR: indicates interquartile range; GLCM: gray-level co-occurrence matrix; GLRL: gray-level run length; SRLGE: short-run low gray-level emphasis; SRHGE: short-run high gray-level emphasis; GLGM: gray-level gradient matrix; SRE: short-run emphasis; LRE: long-run emphasis; GLN: gray-level nonuniformity; RLN: run-length nonuniformity; RP: run percentage; LGRE: low gray-level run emphasis; HGRE: high gray-level run emphasis; SRLGE: short-run low gray-level emphasis; SRHGE: short-run high gray-level emphasis; LRLGE: long-run low graylevel emphasis; LRHGE: long-run high gray-level emphasis; MGR: mean gradients; VGR: variance of gradients. Bold indicates statistically significant as determined with the two-tailed t test and false detection analyses (Q < 0.05).
T A B L E 2 Texture parameters: mean by flip angle.

3.B | Assessment of flip angle
Variations in flip angles produced variations in texture analysis features, as shown in Table 2 and Table S1. Only two histogram fea-

3.C | Assessment of NEX
Changes in NEX (1 vs 4) produced variations in texture analysis features as shown in Table 3

3.D | Assessment of scanner platform
Differences in scanner platform (GE vs Siemens) produced differences in the texture analysis features as shown in Table 4  Mean texture analysis features variation with changes in flip angle. n: number of contoured slices; STD: standard deviation; STD5: 5-neighborhood standard deviation; STD9: 9-neighborhood standard deviation; IQR: indicates interquartile range; GLCM: gray-level co-occurrence matrix; GLRL: gray-level run length; SRLGE: short-run low gray-level emphasis; SRHGE: short-run high gray-level emphasis; GLGM: gray-level gradient matrix; SRE: short-run emphasis; LRE: long-run emphasis; GLN: gray-level nonuniformity; RLN: run-length nonuniformity; RP: run percentage; LGRE: low gray-level run emphasis; HGRE: high gray-level run emphasis; SRLGE: short-run low gray-level emphasis; SRHGE: short-run high gray-level emphasis; LRLGE: long-run low gray-level emphasis; LRHGE: long-run high gray-level emphasis; MGR: mean gradients; VGR: variance of gradients. Bold indicates statistically significant as determined with the two-tailed t-test and false detection analyses (Q < 0.05).
T A B L E 3 Texture parameters: number of excitations 1 vs 4.  There are several limitations to the current study. The first is that this was a study using a nonanatomic phantom with basic architecture variations in internal structure. The use of this phantom and associated scanner data was advantageous as an initial pilot investigation into the dependency texture features on MRI scanning parameters as the raw scanning data are publicly available for research efforts. The phantom used in this study has a well-defined, well-characterized, and simple internal geometric structure. We recognize that the simplicity of this phantom is a far reach from a phantom with anatomically relevant internal structure, but we feel that the simplicity of this nonanatomic phantom initially helps us to understand the results of this study and the effects the changes in MRI scanning parameters has on the texture features. Future research efforts will need to be conducted using a phantom with more anatomically relevant internal structure and with more complex internal components, perhaps with an internal composi-

| CONCLUSION
Texture analysis represents an increasingly popular, post-processing, quantitative evaluation technique that can potentially be used as an adjunct in diagnostic imaging, and as a possible imaging biomarker.
The results of this study demonstrate that MRI acquisition parameters have a significant influence on specific texture analysis features. This work serves as a pilot study highlighting the importance of using a standardized and controlled MRI scanning protocol when using a texture analysis. Multi-institutional research endeavors, or single institution endeavors using different MRI scanning platforms and scanning protocols should exercise caution when using texture analysis.