Radiomic analysis of magnetic resonance fingerprinting in adult brain tumors

Abstract

PURPOSE:

This is a radiomics study investigating the ability of texture analysis of MRF maps to improve differentiation between intra-axial adult brain tumors and to predict survival in the glioblastoma cohort.

METHODS:

Magnetic resonance fingerprinting (MRF) acquisition was performed on 31 patients across 3 groups; 17 glioblastomas, 6 low-grade gliomas, and 8 metastases. Using regions of interest for the solid tumor and peritumoral white matter on T1 and T2 maps, second-order texture features were calculated from gray-level co-occurrence matrices and gray-level run length matrices. Selected features were compared across the three tumor groups using Wilcoxon rank-sum test. Receiver operating characteristic curve analysis was performed for each feature. Kaplan-Meier method was used for survival analysis with log rank tests.

RESULTS:

Low-grade gliomas and glioblastomas had significantly higher run percentage, run entropy, and information measure of correlation 1 on T1 than metastases (p<0.017). The best separation of all three tumor types was seen utilizing inverse difference normalized and homogeneity values for peritumoral white matter in both T1 and T2 maps (p<0.017). In solid tumor T2 maps, lower values in entropy and higher values of maximum probability and high gray run emphasis were associated with longer survival in glioblastoma patients (p<0.05). Several texture features were associated with longer survival in glioblastoma patients on peritumoral white matter T1 maps (p<0.05).

CONCLUSION:

Texture analysis of MRF-derived maps can improve our ability to differentiate common adult brain tumors by characterizing tumor heterogeneity, and may have a role in predicting outcomes in patients with glioblastoma.

INTRODUCTION

Tumors with high intratumoral heterogeneity typically have worse prognosis likely secondary to their intrinsically aggressive biology [1]. Intra-tumoral heterogeneity in malignant brain tumors, particularly in gliomas, occurs at cellular as well as molecular levels [2]. On histopathologic analysis, areas of local and regional heterogeneity are nearly always identified. For example, areas of lower grade neoplasm are often interspersed within high-grade gliomas [1,2]. At a molecular level, gliomas can be classified on the basis of somatic mutations in isocitrate dehydrogenase (IDH) type I/II, TP53; transcriptional signature, copy number variations such as co-deletion of 1p and 19q and epidermal growth factor (EGFR) mutation. Several of these subtypes have been associated with distinct patterns of therapeutic response and outcomes [2–5]. Tumor heterogeneity also plays an important role in treatment response and clinical outcomes in patients with lower grade gliomas as well as metastases [6–8]. Single specimen tissue biopsies are usually inadequate to capture the tumor heterogeneity, and even subtotal resections can fail to fully represent intra-tumoral heterogeneity [1,2].

In contrast to focal biopsies, imaging has the potential to capture local and regional heterogeneity, particularly using a radiomics approach. Radiomics holds promise as a tool in the diagnosis of various cancer types, treatment planning, and response prediction while assessing associations with patient outcomes [9–18]. This method quantifies the spatial distribution of pixel intensities in an image in the form of quantitative texture features which are not discernable to the unaided human eye [19,20]. Although, numerous approaches have been proposed to quantify and characterize image texture; the most commonly used methods are implemented by deriving first- or second- order statistical features [20–22]. Magnetic resonance (MR) imaging texture analysis has been used to analyze glioblastoma data with several texture features correlating with survival outcomes [3,23–27]. Radiomics has also been used in lower grade gliomas to non-invasively estimate IDH1 status [28,29]. Traditionally, qualitative MR images served as the basis for texture analysis but recently quantitative MR imaging techniques have been used with this analysis, including magnetic resonance fingerprinting (MRF) [30–33].

MRF is a rapid imaging technique in which tissue properties such as T1 and T2 relaxation times can be calculated simultaneously in a single acquisition [32,34]. In our previous work, we evaluated the ability of MRF-derived first order statistical features to differentiate 3 common types of intra-axial brain tumors including glioblastoma, low grade gliomas (LGG) and metastases [35]. In the current study, we hypothesized that texture analysis of 2D MRF maps will improve our ability to capture distinct phenotypic characteristics of these tumors and enhance tumor differentiation. We also assessed if there was any correlation between various texture features and overall survival in the glioblastoma cohort.

MATERIAL AND METHODS

This study was approved by the University Hospitals Cleveland Medical Center institutional review board to evaluate clinical applications of MRF. Written informed consent was obtained from all subjects.

The inclusion criterion encompassed adult patients with untreated intra-axial neoplasms at the time of the scan. The exclusion criteria comprised of all contraindications to MRI including patients with implanted pacemakers, metallic implants, and a history of claustrophobia. Additional exclusion criteria included contraindications to gadolinium contrast such as pregnancy and prior allergic reactions to MR contrast material [35]. Figure 1 gives an overview of the data processing and analysis workflow.

Fig I.

Schematic diagram for data processing and analysis.

Subjects:

As described in our previous work, seventeen patients with glioblastomas, eight with brain metastases, and six with WHO grade II glial neoplasms including five oligodendrogliomas and one oligoastrocytoma were included in the study. Final histopathological diagnosis was available in all patients (13 total resections, 16 partial resections and 2 biopsies). Table I outlines patient demographics, including age and sex.

Table I.

Patient demographics of 31 patients with untreated brain neoplasms [35]

	GBM (n=17)	LGG (n=6)	Metastasis (n=8)	P Value
Age, yr (mean) (range)	61.4 ± 9.2 (45–76)	46.5 ± 12.1 (38–67)	63.5 ± 8.6 (48–76)	0.03a
Sex (No.)				0.57b
Female	8 (47.1%)	2 (33.3%)	5 (62.5%)
Male	9 (52.9%)	4 (66.7%)	3 (37.5%)
Steroids (No.)c				0.77b
Yes	5 (29.4%)	1 (16.7%)	3 (37.5%)
No	12 (70.6%)	5 (83.3%)	5 (62.5%)
IDH1 (No.)				0.003d
Positive	0 (0.0%)	4 (66.7%)	NA
Negative	11 (100.0%)	1 (16.7%)e	NA

Note: - NA indicates not applicable.

^aP value from the Kruskal-Wallis test. Results of the Wilcoxon rank sum tests showed that the LGG group differed in age from the GBM and metastasis groups (P = 0.014 and 0.023, respectively) and that the GBM and metastasis groups did not differ in age (P = 0.68).

^bP value from an exact version of the Pearson χ² test comparing proportions in the 3 groups.

^cWilcoxon rank sum test revealed no differences in T1 and T2 values when patients with the presence and absence of steroid treatment were compared by tumor type.

^dP value from the Fisher exact test comparing GBM and LGG groups.

^eIDH1 status of 1 patient with LGG was unknown.

MRF Acquisition:

All patients were scanned at 3T (Verio and Magnetom Skyra; Siemens, Erlangen, Germany) using a 20-channel head coil. MRF acquisitions were obtained for all patients immediately prior to the clinical MR scan with single slice acquisition through the most representative area of the tumor. The MRF acquisition consisted of a True FISP sequence with the following parameters; FOV, 300 × 300 mm²; matrix, 256 × 256; section thickness, 5 mm; flip angle variable, 0°–60°; TR variable, 8.7–11.6 ms; sinc radiofrequency pulse with a duration of 800 s; and a time-bandwidth product of 2. 3000 time points were acquired with a total acquisition time of 30.8 seconds for every section. The final output included a T1, T2, proton-density, and off-resonance map. Our quantitative analysis utilized only the T1 and T2 maps [35].

ROI Processing:

A fellowship-trained neuroradiologist and a board-certified radiologist who were both blinded to the final pathology drew ROIs on the quantitative T1 and T2 maps utilizing axial FLAIR and postcontrast T1-weighted images for reference. ROIs for solid tumor (ST) and peritumoral white matter (PW) regions were drawn for each tumor. The ST region was defined as the enhancing region in tumors with postcontrast enhancement or an expansile FLAIR hyperintensity region in nonenhancing/minimally enhancing tumors. The PW region was defined as white matter within 1 cm of the enhancing or expansile FLAIR hyperintense tumor margin. The size of the lesions and consequently the ROIs ranged from 0.32 to 12 cm² (median, 1.7 cm²) for ST regions and from 0.25 to 2.5 cm² (median, 0.96 cm²) for PW regions. ROIs drawn by both readers were compared by means of intraclass correlation coefficient as an assessment for interobserver concordance and the set of ROIs drawn by the neuroradiologist were used for further analysis [35]. After each ROI was drawn, the rest of the image data was removed. The resulting ROI region was converted to gray scale before each pixel was binned into 1 of 32 intensities. Typically, the maximum number of gray levels considered for each image or ROI is scaled down to 32 or 64 gray levels which allows robust texture analysis without masking of important image features. [21,36,37].

Second Order Texture Features:

Quantitative second order textural features were computed through gray level co-occurrence matrix (GLCM) and gray level run length matrix (GLRLM). GLCM is used to analyze local tumor features by tabulating the frequency in which one-pixel intensity appears directly adjacent to any other pixel intensity within an image. This joint probability of the intensity values of two pixels being directly adjacent takes the form of a square matrix with row and column dimensions equal to the number of discrete gray level intensities in the image [20,22,38,39]. We created four 32 × 32 GLCMs with angular directions of 0, 45, 90 and 135. Ultimately, resulting second order texture features were averaged across these four directions. Based on the literature review, 18 second-order features were computed by means of GLCMs for both ST and PW region (Table II) [3,22,36,40–42]. GLRLM is used to analyze whole tumor features by tabulating the run length of a given intensity throughout the entirety of the tumor. The run lengths are organized into a 32 × 256 matrix to accommodate the max possible run length for each of the 32 discrete gray levels. Similar to GLCM, 4 matrices are created, one for each of the angular directions of 0, 45, 90, and 135. Using this approach 16 GLRLM features were extracted (Table II) [3,22,40]. Figure II shows the example images of MRF-derived quantitative T1 and T2 maps, the appearance of the same lesions on FLAIR and post-contrast T1-weighted images, and ROI delineation and conversion of the ROIs to the GLCM and GLRLM in one of the study participants. All analyses were performed using MatLab R2017b (MathWorks, Natick, Massachusetts).

Table II.

List of texture features and used abbreviations.

GLCM Features
Autocorrelation*
Cluster-prominence*
Cluster-shade*
Contrast*
Correlation*
Difference entropy
Difference variance
Dissimilarity
Energy
Entropy*
Homogeneity*
Information measure of correlation 1 (IMC1) *
Information measure of correlation 2 (IMC2)
Inverse difference
Inverse difference moment normalized
Inverse difference normalized (IDN)*
Inverse variance*
Joint average
Maximum probability*
Sum average
Sum entropy
Sum of squares variance
Sum variance
GLRLM Features
Short run emphasis*
Long run emphasis
Gray-level nonuniformity (GLN)*
Gray-level nonuniformity normalized
Run length nonuniformity*
Run length nonuniformity normalized (RLNN)*
Run percentage (RP)*
Gray level variance
Run variance
Run entropy (RE)*
Low gray-level run emphasis*
High gray level run emphasis*
Short run low gray-level run emphasis
Short run high gray-level emphasis
Long run low gray-level emphasis
Long run high gray-level emphasis*

Features with asterisk (*) were selected for further analysis.

Fig II.

a: T2-weighted image of a 39-year-old male who was diagnosed with left frontal oligodendroglioma b-c: FLAIR and T1 post-contrast images from the clinical scan d-e: Magnetic Resonance Fingerprinting-derived quantitative T1 and T2 maps with ROI overlay f: Conversion of the region of interest to the gray level scales on the gray level co-occurrence matrix

Feature Selection and evaluation:

As a first step, Spearman’s rank correlation coefficient with a cut-off value of 0.85 and p-value <0.05 was used to identify highly correlating features. One feature from each pair was selected to avoid duplication and the remainder was removed from further analysis. The Wilcoxon rank sum test was used to compare each selected texture feature across three tumor groups. The result was considered significant if p-value was < 0.017 after applying the Bonferroni method for multiple comparison correction. Receiver operating characteristic (ROC) curve analysis was utilized to evaluate the ability of texture features to discriminate between tumor types. Area under the curve (AUC) was calculated for each feature. For survival analysis, overall survival (OS) was defined as the number of months from the date of diagnosis of glioblastoma that the patients were known to be alive. In our cohort of 17 patients with glioblastomas, two were alive on the date of the analysis. To evaluate the relationship between texture features and OS, we divided the patients into two subgroups using the median value of each feature as cut-off for each region (ST or PW) and each map (T1 or T2). The patient with the median value for each feature was included in the group with values above the cut-off creating two equally sized groups for comparison. For each feature, a log-rank test was used to measure the differences between the survivals of subjects in the two subgroups. Survival results with p -value < 0.05 were considered as statistically significant. All statistical analyses were performed using SPSS V26, (SPSS Inc, Chicago, Illinois).

RESULTS

Based on the Spearman’s rank correlation coefficient test results, 20 out of 39 texture features were selected for further analysis (Table II). The Wilcoxon rank sum test for feature comparison across ST regions of tumor groups reveal no difference in texture features between glioblastomas and LGGs for T1 and T2 maps. Metastases have lower IMC1, RE, and RP compared to LGGs (p=0.005, p=0.005, and p=0.013 respectively) and glioblastomas (p=0.001, p=0.002, and p=0.006 respectively) on T1 maps. Similarly, metastases have lower GLN, RE, and RP than LGGs (p=0.005, p=0.005, and p=0.013) on the T2 maps. In contrast, RLNN is higher in metastases than LGGs (p=0.005).

In the analysis of PW regions, multiple texture features were able to differentiate between metastases, LGGs and glioblastomas. Homogeneity and IDN have lower T1 values in glioblastomas compared to metastases (p=0.009 and p=0.003 respectively). Metastases also have higher T1 and T2 contrast than glioblastomas (p=0.004 and p=0.013). LGGs had similar trends in differentiating from metastases in the T1 and T2 maps (p<0.05). T2 homogeneity and IDN can be used to differentiate between glioblastomas and LGGs (p=0.013 and p=0.010 respectively). Table III illustrates the areas of significance for each of the comparisons being done across T1 and T2 maps in addition to ST and PW regions.

Table III.

Wilcoxon rank-sum test p values for tumor comparison among solid tumor (ST) and peritumoral white matter (PW) in T1 and T2 maps for 20 texture features. Significant results after the Bonferroni correction (p<0.017) are highlighted in red. Non-significant results (p>0.017) are highlighted in green

	AutoCorrelation	ClusterProminence	ClusterShade	Contrast	Correlation	Entropy	GLN	HGRE	Homogeneity	IMC1	InverseDifferenceNormalized	InverseVariance	LGRE	LRHGE	Maximum Probability	RE	RLN	RLNN	RP	SRE
GM vs LGG	0.759	1.000	0.919	0.286	0.354	0.473	0.286	0.562	0.759	0.919	0.392	0.431	0.919	1.000	0.759	0.256	0.658	0.227	0.256	0.973	ST_T1
LGG vs Mets	0.852	0.662	0.950	0.029	0.008	0.081	0.026	0.240	0.059	0.005	0.029	0.029	0.240	0.937	0.852	0.005	0.132	0.041	0.013	0.818
GM vs Mets	0.932	0.549	0.406	0.124	0.262	0.238	0.024	0.256	0.013	0.001	0.049	0.086	0.354	0.812	0.669	0.002	0.135	0.201	0.006	0.865
GM vs LGG	0.562	0.227	0.087	0.201	0.407	0.919	0.087	1.000	0.658	0.802	0.392	0.812	0.609	0.473	0.759	0.117	0.473	0.117	0.117	0.473	ST_T2
LGG vs Mets	0.662	0.142	0.282	0.020	0.043	0.282	0.005	0.138	0.181	0.345	0.059	0.228	0.295	0.731	0.662	0.005	0.101	0.005	0.013	0.366
GM vs Mets	0.110	0.194	0.628	0.066	0.238	0.315	0.034	0.087	0.140	0.291	0.075	0.157	0.534	0.209	0.842	0.016	0.187	0.187	0.057	0.901
GM vs LGG	0.708	0.812	0.708	0.392	0.562	0.562	0.831	0.521	0.286	0.812	0.177	0.319	0.416	0.639	0.392	0.256	0.765	0.898	0.473	0.467	PW_T1
LGG vs Mets	0.491	0.282	0.852	0.005	0.059	0.852	0.082	0.052	0.013	0.414	0.003	0.059	0.082	0.126	0.029	0.013	0.052	0.537	0.005	0.662
GM vs Mets	0.549	0.140	0.215	0.004	0.043	0.406	0.046	0.075	0.009	0.238	0.003	0.194	0.173	0.633	0.049	0.023	0.014	0.289	0.049	0.775
GM vs LGG	0.155	0.516	0.708	0.135	0.227	0.020	0.718	0.051	0.013	0.658	0.010	0.919	0.091	0.033	0.062	0.201	0.062	0.312	0.101	0.051	PW_T2
LGG vs Mets	0.108	0.282	0.755	0.008	0.059	0.081	0.132	0.026	0.003	0.345	0.003	0.852	0.180	0.065	0.081	0.020	0.065	0.310	0.059	0.132
GM vs Mets	0.887	0.238	0.842	0.013	0.049	0.887	0.130	0.904	0.194	0.374	0.027	0.798	0.659	0.659	0.588	0.086	0.659	0.274	0.511	1.000

On ROC analysis of the PW regions, entropy values on T2 maps offers the best separation between glioblastomas and LGGs with an AUC of 0.869 (95% confidence interval 0.71–1.00, p=0.011). T2 GLN for ST regions give the best separation of LGGs and metastases with an AUC of 0.952 (95% confidence interval 0.84–1.00, p=0.007). Correlation values on T1 maps for PW regions give the best separation of glioblastomas and metastases with an AUC of 0.877 (95% confidence interval 0.71–1.00, p=0.016). Figure III summarizes the significant results in the form of box plots.

Fig III.

Boxplots representing the distribution of significantly different features across various tumor groups a: ST region IMC1 values on T1 maps b: ST region run percentage values on T1 maps c: ST region run entropy values on T2 maps d: PW region homogeneity values on T1 maps e: PW region contrast values on T2 maps f: PW region homogeneity values on T2 maps

In the survival analysis of glioblastoma patients, there was a significant difference between survival time of patients with lower T2 entropy values in ST regions as compared to patients with higher entropy values (p=0.034) with median survival estimate of 11 months for patients below the cut-off value and 6.7 months for patients above the cut-off value. Meanwhile, higher T1 entropy values in PW regions are correlated to higher survival (p=0.009) with a median survival estimate of 6.8 months for patients below the cut-off value and 18 months for patients above the cut-off value. Table IV highlights other texture features that are strongly correlated to overall survival for glioblastoma patients. Figure IV summarizes the significant results in the form of Kaplan-Meyer survival curves.

Table IV.

Kaplan-Meier analysis p-values of comparing texture features above and below the median. Significant results (p<0.05) are highlighted in red. Non-significant results (p>0.05) are highlighted in green.

	AutoCorrelation	ClusterProminence	ClusterShade	Contrast	Correlation	Entropy	GLN	HGRE	Homogeneity	IMC1	InverseDifferenceNormalized	InverseVariance	LGRE	LRHGE	Maximum Probability	RE	RLN	RLNN	RP	SRE
ST_T1	0.123	0.630	0.847	0.531	0.163	0.211	0.923	0.810	0.178	0.962	0.360	0.596	0.810	0.054	0.336	0.060	0.211	0.700	0.336	0.563
ST_T2	0.470	0.312	0.531	0.470	0.360	0.034↓	0.923	0.043↑	0.178	0.630	0.736	0.123	0.336	0.194	0.009↑	0.700	0.067	0.700	0.810	0.596
PW_T1	0.665	0.312	0.312	0.360	0.962	0.009↑	0.884	0.006↓	0.386	0.847	0.773	0.014↑	0.011↑	0.011↓	0.012↓	0.962	0.025↓	0.884	0.123	0.261
PW_T2	0.470	0.312	0.885	0.048↓	0.163	0.885	0.229	0.329	0.810	0.736	0.290	0.229	0.596	0.305	0.149	0.962	0.665	0.773	0.885	0.312

↑ indicates higher values correlate to higher survival

↓ indicates lower values correlate to higher survival

Fig IV.

Kaplan-Meier survival curves in the glioblastoma cohort using the median feature values as cut-off. a: Survival curves for the groups with T2 Entropy values in solid tumor regions below and above the cut-off value. b: Survival curves for the groups with T1 Entropy values in peritumoral white matter regions below and above the cut-off value.

Interobserver concordance was determined to be excellent with intraclass correlation coefficients for ST T1, ST T2, PW T1, and PW T2 of 0.90, 0.83, 0.88, and 0.89 respectively [35].

DISCUSSION

This study describes application of texture analysis to MRF data on malignant brain tumors. Texture analysis is an established powerful tool to indirectly capture local and regional heterogeneity in tumors and when applied to MRF can identify significant differences in various tumor types. Certain MRF derived texture features demonstrate potential for stratification of glioblastoma patients based on overall survival.

Spatial variation in tissue cellularity, presence of necrosis, degree of angiogenesis and amount of extravascular-extracellular matrix are the most well studied histopathological determinants of tumor heterogeneity [43–47]. T1 and T2 relaxation times have also shown to be affected by water content, cellularity, cell type, macromolecular content, presence of paramagnetic substances and tissue architecture [48]. On the other hand, relaxometry differences could conceivably be masked if tumors have similar cellular origins or morphology. The radiomics approach can overcome this limitation by identifying textural patterns that are unique to each tumor type. This was demonstrated in our study in which we were able to identify textural differences between solid tumor and peritumoral regions of glioblastomas and metastases, which were not identified in our initial work [35].

Within a region of interest, contrast measures the difference in intensity between pixels with larger values of contrast indicating a larger difference between surrounding pixels. IMC1 and RE are functions of entropy, which in turn are a measure of randomness of different gray level values. Our results show higher IMC1 and RE for glioblastomas and LGGs relative to metastases which is consistent with higher tissue heterogeneity in glioblastomas and LGGs. Homogeneity is a measure of gray level similarity between pixel pairs while IDN is another measure of homogeneity; lower values indicate greater heterogeneity [20,38,49]. The lower levels of homogeneity and inverse difference normalized for glioblastomas indicate greater peritumoral white matter heterogeneity than LGGs. Being able to demonstrate these differences quantitatively is valuable in deepening our understanding of glioma biology. The peritumoral region surrounding glioblastoma consists of edema combined with varying amount of tumor infiltration. This is in contrast to the ‘pure’ vasogenic edema in the peritumoral regional around metastases. These differences are not appreciated on clinical MRI images. Our approach demonstrates quantitative differences between peritumoral regional of glioblastomas and metastases with lower homogeneity and IDN in former as compared to latter. This finding may have implications in tumor delineation and treatment planning in the future.

Several studies have focused relationship between texture features and clinical outcomes in brain tumor patients [3,23,24]. Using 3D heterogeneity measures derived from post contrast T1 weighted images, Molina et al. demonstrated that low entropy and high homogeneity were associated with improved survival [3]. Chaddad et al. used texture features derived from FLAIR and post contrast T1 images and found that features such as energy, correlation, and variance were useful for predicting patient survival time [23]. A more recent study by Liu et al comprehensively evaluated various features derived from co-occurrence matrix, run-length matrix and histogram analysis and found that features measuring local and regional heterogeneity play an important role in survival stratification [24]. In our study, lower T1 entropy meaning lower heterogeneity in the ST region was significantly correlated with longer survival. Analysis of peritumoral white matter offers additional features that strongly correlate with survival. This offers a chance to predict clinical outcomes in brain tumor patients based on analysis of the baseline scans. Our study is fundamentally distinct from the previous approaches as the texture features have been derived entirely from quantitative maps. Given the quantitative nature of the raw data, our radiomics approach may have a distinct advantage as compared to the conventional radiomics approach, particularly in ensuring feature robustness and generalizability [33,50].

Image based radiomics analysis has been used in several studies, however, so far, there is no consensus on an optimal feature extraction or selection method [3,21,24,25,51]. The 39 features that we initially selected were based on commonly used features in previous literature using qualitative MRI images. In addition to GLCM and GLRLM analyses, texture features computed through Gray Level Size Zone Matrix (GLSZM), Neighboring Gray Tone Difference Matrix (NGTDM) and Gray Level Dependence Matrix (GLDM) can also be used for such analyses, however these texture features were not included in our preliminary proof-of concept study design. To our knowledge, there are few texture analysis studies based on quantitative data [30,31,33]. To improve the robustness of our results, we excluded the highly correlating features using Spearman’s rank correlation and applied multiple comparisons testing correction.

A key limitation of this work is single slice two-dimensional acquisition of MRF data which does not provide whole tumor coverage. Future studies with 3D volumetric MRF acquisition and targeted histologic correlation are essential to map the solid and peritumoral regions in their entirety. Given the modest sample size and the large number of statistical tests performed in this study, the results need validation with larger sample studies with appropriate multiple comparison correction. The relationship between MRF texture features and patient outcomes must be interpreted cautiously given the small sample size. Further assessment with larger datasets using multiple regression models will assess the strength of this relationship while adjusting for known predictors of outcomes. If these results are validated, MRF radiomic analysis may have an important role in management of brain tumors.

CONCLUSION

This study demonstrates the application of texture analysis on MRF-derived quantitative maps in evaluation of common malignant brain tumors. The results suggest that texture analysis offers a unique method of analysis of MRF data that allows tumor differentiation above and beyond the first order features and may have role in predicting overall survival in patients with glioblastoma.

ABBREVATIONS

AUC

Area under the curve

GLCM

Gray level co-occurrence matrix

IMC1

Information measure of correlation 1

LGG

Lower grade glioma

MR

Magnetic resonance

MRF

Magnetic resonance fingerprinting

OS

Overall survival

ROC

Receiver operating characteristic

ROI

Region of interest

ST

Solid tumor

PW

Peritumoral white matter

Availability of data and material:

Upon reasonable request

Code availability:

NA