The utility of multi-b-value diffusion and arterial spin labelling magnetic resonance imaging in gliomas grading and prediction of isocitrate dehydrogenase status

Introduction

Differentiation between high-grade glioma (HGG) and low-grade glioma (LGG) by magnetic resonance imaging (MRI) is essential for choosing adequate therapy and for estimating a patient’s prognosis. However, the possibilities of conventional MRI methods in this application are somewhat limited. Although contrast enhancement of brain gliomas may be regarded as a sign of malignancy, in approximately one-third of cases contrast enhancement is not present (1). This fact may also be important in the follow-up of patients with surgically treated LGG, where malignant transformation of residual tumour may occur quite frequently and its detection is important for the choice of further treatment strategy (2). For these reasons, considerable effort has recently been devoted to the development of non-invasive advanced magnetic resonance (MR) imaging techniques useful for determining the grade of gliomas. One of the approaches intensively investigated consists in various techniques of MR perfusion imaging, whether these be methods using contrast agent application, such as dynamic susceptibility contrast (DSC) or dynamic contrast enhancement (DCE), or the non-contrast MR perfusion technique arterial spin labelling (ASL). Using these techniques, strong correlation of tumour vascularity with histological grade or molecular status of isocitrate dehydrogenase (IDH) gene mutation has been found (3,4). Another technique that is widely used to determine the degree of malignancy in gliomas is the diffusion-weighted imaging (DWI) method. A basic methodological approach can be taken that uses a monoexponential diffusivity model with determination of apparent diffusion coefficient (ADC) values. The application of this method is based on an assumption that highly cellular malignant tumours have relatively limited diffusivity, as has been demonstrated by several studies using the ADC parameter for grading gliomas (5). More advanced diffusion imaging data acquisition and analysis models, such as diffusion tensor imaging (DTI) or diffusion kurtosis imaging (DKI), also have been used for glioma grading, offering additional insight into the microstructure of the tumour tissue under investigation and using quantification with new parameters reflecting also diffusivity anisotropy (6,7). Another interesting approach to diffusion data analysis is intravoxel incoherent motion (IVIM). This technique, based on a biexponential model for the analysis of diffusion data measured with multiple b-factor values (multi-b), also reflects to some extent the tissue perfusion, where IVIM parameters have been shown to correlate with those of conventional MR perfusion imaging techniques such as DSC (8). Some attempts already have been made to use IVIM for glioma grading and achieved quite promising results, with perfusion fraction (f) and pseudo-diffusion (D*) parameters seeming to correlate with the histopathological grade (9,10). Limited number of studies comparing IVIM and ASL techniques suggest a higher diagnostic yield of perfusion parameters derived from diffusion data using the IVIM model in comparison to cerebral blood flow (CBF) values obtained through the ASL technique (10,11). However, there is relatively little work on this topic to date. The methods of the diffusion parameters analysis are various and the results not entirely uniform (12). Therefore, we believe this topic deserves further exploration. Furthermore, by including images with higher diffusion weighting, it is possible to characterize also the non-Gaussian behaviour of water diffusion through diffusion kurtosis, Thus, this study presents a comprehensive diffusion MRI technique combining IVIM and non-tensorial DKI that, in conjunction with three-dimensional (3D) segmentation and histogram analysis, offer the potential for an evaluation of tumour diffusivity and perfusion characteristics. Compared to complex multi-shell diffusion sequences, this approach benefits from a shorter acquisition time, while the isotropic data obtained allow for the calculation of a range of potentially valuable parameters. The main aim of this work is to evaluate their potential for classifying brain gliomas according to their histopathological grade and IDH status and to compare the diagnostic yield of diffusion parameters with the better-established native perfusion imaging technique of ASL. We present this article in accordance with the STARD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1182/rc).

Methods

Patient selection

The prospective single-centre study was approved by the Ethics Committee of the University Hospital Brno (No. 21-100620/EK), written informed consent was obtained from all participants. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. We enrolled 326 consecutive patients scheduled for surgery within the years 2019–2023 for intracranial brain lesion with suspicion of brain tumour according to the previous imaging performed. Patients with prior history of brain tumour surgery were not included. For this study, we excluded patients whose biopsy result was other than glioma or who finally were not operated, as well as patients with significant motion artifacts on MRI or those whose examinations were incomplete. In the end, 107 patients with histologically confirmed gliomas entered the MRI data analysis (Figure 1).

Figure 1 Flow diagram of the selection process. MRI, magnetic resonance imaging.

MRI acquisition

The examinations were performed on a 1.5T MR device (Philips Ingenia, Best, Netherlands) with 20 channel head-neck coil and gradient system with maximum amplitude 33 mT/m and slew rate 120 T/m/s. The imaging protocol comprised DWI single-shot echo planar imaging (SS-EPI) sequence in axial plane using 10 different b values (0, 10, 20, 30, 50, 100, 200, 500, 1,000, 2,000) and sequences for morphological evaluation, including non-contrast enhanced (CE) two-dimensional (2D) T1 fast field echo (FFE), 2D T2 turbo spin echo (TSE) in axial plane, CE 3D T1 TSE acquisition in axial plane, and 3D fluid attenuated inversion recovery (FLAIR) in sagittal plane. The same SS-EPI DWI sequence with only b=0 and opposite phase-encoding direction was acquired for susceptibility artefact correction. More detailed description of diffusion MRI sequence is shown in Table 1 and further details about morphological MRI sequences are shown in electronic supplementary material (Table S1).

For ASL acquisition, 3D GRASE sequence was acquired [repetition time (TR) =4,000 ms, echo time (TE) =13 ms] in normalized pseudo-continuous regime [pseudo-continuous ASL (pCASL)] as 9 dynamic scans with post label delay 1,800 ms. Single 3D block covered whole brain with slice thickness 8 mm and in-plane resolution 3.75×3.75 mm².

Morphological segmentation

All operations with the images were done by scripts from FSL library (14). All morphological images (T2, FLAIR, and T1) were registered into T1-CE space by linear rigid (6 degrees of freedom) algorithm (FLIRT) and skull striped by bet (15,16). All segmentations were performed by two board-certified radiologists with more than 6 years’ experience (T.K. and T.J.) using a semi-automatic “classification” algorithm based on decision trees and active contour evolution implemented in ITK-SNAP 3.8.0 software (17). The segmentation mask entering further processing covered both enhancing and non-enhancing parts of the tumour and the zone of surrounding oedema, if present; therefore, any necrotic areas were not excluded from the final volume of interest. All results were visually checked in ITK-SNAP interface and manually corrected when needed.

Multi-b diffusion and ASL processing

Several pre-processing steps were made using scripts from FSL library. Susceptibility distortion correction was done by the “topup” algorithm followed by eddy current and movement artifacts correction by the “eddy” algorithm (18,19). Scull striping was done by “bet” algorithm and anatomical T2-weighted image was registered into b=0 space by linear rigid (6 degrees of freedom) registration using the “FLIRT” algorithm. The binary masks of segmented tumours were registered into the DWI space by the same registration matrix as used for the T2-weighted image in the previous step. The first multi-b diffusion analysis was based on the IVIM model. For fitting of IVIM parameters [f, D*, and D (diffusion)], algorithms from the Diffusion in Python (DIPY) (20) library were used together with the optimization process method known as “variable projection”, which utilizes an MIX approach with bounds 0–0.8 for f, 0–0.1 mm²/s for D*, and 0–0.01 mm²/s for D and all available b-values were used (21). Additional multi-b diffusion analysis was based on a diffusion kurtosis model wherein parameters of diffusion and kurtosis (K) were fitted by logarithmic fitting option in Medical Imaging Interaction Toolkit (MITK) software with boundaries from 0 to 3 and smoothing sigma equal to 5 and all available b-values were used (22,23). For the subsequent statistics, only the diffusion parameter (D) calculated by IVIM model was used and not the one calculated by the diffusion kurtosis model. Examples of multi-b parameter maps and 3D volumes of interests (VOIs) in patients with gliomas are presented in Figures 2-4.ASL normalization by proton density-weighted images as well as CBF maps were calculated automatically by hardware-related Philips software. ASL image was subsequently registered to DWI b=0 space by linear rigid (6 degrees of freedom) registration using the “FLIRT” algorithm. For the final analyses, the same binary segmentation masks as in multi-b analyses were used.

Figure 2 Example of the segmentation masks in patients with LGG and HGG. Examples of the segmentation masks in patients with low-grade glioma (LGG, upper row) and with high-grade glioma (HGG, lower row) shown in axial (A,E), sagittal (B,F) and coronal (C,G) reconstructions of T1-weighted contrast-enhanced images and as 3D reconstructions (D,H). Patients and axial scans are the same as in Figures 3,4. The blue lines in the image correspond to the positions of the images in different planes. Voxels from these 3D volumes served as input data for all analytical steps. 3D, three-dimensional; HGG, high-grade glioma; LGG, low-grade glioma.

Figure 3 Multi-b and ASL MRI examination in a patient with HGG. MRI examination in 62-year-old women with HGG (glioblastoma grade 4, IDH wild-type) located in medial rostral part of the right occipital lobe depicted as an enhancing mass in T1 weighted axial image (A). Maps of D—diffusion coefficient (B), D*—pseudo-diffusion (C), f—perfusion fraction (D), FLAIR image (E), CBF (f), fD* (G) and K (H) are shown. High signal intensity in FLAIR image and hyper-vascularized rim is well visible in ASL (CBF) image as well as in D* and fD* maps. Comparatively high values were observed also for f and K parameters. Map of D shows a spot of restricted diffusivity. Mean K and D* values measured within the tumour were 0.834 and 10,465 (10−6 mm²/s), respectively. ASL, arterial spin labelling; CBF, cerebral blood flow; FLAIR, fluid attenuated inversion recovery; HGG, high-grade glioma; IDH, isocitrate dehydrogenase; MRI, magnetic resonance imaging.

Figure 4 Multi-b and ASL MRI examination in a patient with LGG. MRI examination in a 41-year-old man with diffuse astrocytoma (grade 2, IDH-mutated). The tumour located in the right frontal lobe in axial T1 weighted axial image (A) is hypointense with tiny spots of enhancements in central part. Maps of D—diffusion coefficient (B), D*—pseudo-diffusion (C), f—perfusion fraction (D), FLAIR image (E), CBF (F), fD* (G) and K (H) are shown. High signal intensity is seen in FLAIR image. The diffusivity of the tumour is high, as depicted on D map. Perfusion visible on CBF map is low, as are D*, fD*, and K values, as depicted on parametric maps, and f values are comparatively high. Mean K and D* values measured within the tumour were 0.452 and 4,817 (10⁻⁶ mm²/s), respectively. ASL, arterial spin labelling; CBF, cerebral blood flow; FLAIR, fluid attenuated inversion recovery; IDH, isocitrate dehydrogenase; LGG, low-grade glioma; MRI, magnetic resonance imaging.

Statistical analysis

The subgroups were compared in terms of age and sex distribution by Student’s t-test and Chi-squared test, respectively.

We performed histogram analysis to evaluate all voxels belonging to the segmentation masks, calculating mean; median; standard deviation (SD); minimum; 5th, 25th, 75th, and 95th percentile values; as well as maximum, skewness, and kurtosis for all parameters of diffusion parameters (f, D*, D, fD*, and K) as well as for CBF values derived from ASL data. We also calculated volumes of the segmented tissues. Spearman correlation coefficients between CBF and fD* were calculated for all histogram parameters.

All parameters were compared between subgroups (HGG vs. LGG and IDH-mutated vs. IDH wild-type, respectively) using Mann-Whitney U-test with false discovery rate (FDR) correction for multiple testing. Receiver operating characteristic (ROC) analysis was performed to characterize observed parameters and set optimal thresholds for differentiation between the two subgroups. Odds ratio (OR) and 95% confidence interval were calculated for all thresholds by univariate logistic regression models. Least absolute shrinkage and selection operator (LASSO) regression method was used to regularize and select candidate variables from the multi-b dataset for subsequent multivariable analysis by classification trees. The decision tree was trained using the classification method with 10-fold cross-validation to estimate the model’s performance, which was evaluated based on sensitivity (true positive rate) and specificity (true negative rate). These analyses were done in R software (v4.3.2) using the grpreg v3.4.0, pROC v1.18.5, and rpart v4.1.21 packages. Significance level was set at P<0.05 for all statistical tests.

Results

The final study group included 107 patients with histologically confirmed gliomas, which were graded according to World Health Organization (WHO) classification and the presence of IDH mutation was established in all patients. Figure 1 shows the flow diagram of the patient’s selection process. Three IDH wild-type gliomas initially classified according to previous WHO classification as grade II were reclassified as grade 4 tumours to comply with WHO 2021 classification and thus included into the HGG subgroup. Finally, 83 patients formed HGG group (14 grade 3 and 69 grade 4) and 24 the LGG group (grade 2). Furthermore, the patients were divided according to their IDH status into an IDH-mutated (positive) group of 35 subjects and an IDH wild-type (negative) group of 72 subjects.

Table 2 shows further characteristics of the study group and distribution of the patients according to different histopathological characteristics.

Segmented volumes of tumours were significantly larger in HGG (82±60.8 mL) compared to LGG (45.8±59.9 mL) according to Mann-Whitney U-test (P=0.002). The volumes between IDH positive and negative groups did not differ significantly.

We found a statistically significant correlation between CBF and fD* for Q95 (ρ=0.311, P=0.001) and SD (ρ=0.391, P<0.001) parameters. Histogram parameters min and max were also statistically correlated, but both were affected by the high frequency of zero values in the case of CBF or by the upper limit of the IVIM fit, so these parameters appear not to be reliable.

In analysing histograms of individual diffusion imaging and ASL parameters, we found a total of 49 parameters that differed statistically significantly between subgroups of HGG and LGG patients, and for 48 of these parameters the differences were significant even after correction for multiple testing. Similarly, 42 parameters differed significantly between IDH positive and negative groups (36 after correction for multiple testing). Table 3 shows mean values for all multi-b diffusion parameters and ASL measured within the whole volume of lesions. The full set of histogram parameters can be found in Table S2.

Tables 4,5 show results of ROC analyses and linear regression demonstrating the power of the 10 most significant histogram parameters (sorted by the sum of sensitivity and specificity) to differentiate between LGG and HGG (Table 4) and between IDH-positive and -negative tumours (Table 5).

As for ASL technique, a number of parameters were statistically significantly different between patient subgroups defined by tumour histopathologic grade or the presence of an IDH mutation (Table S2); however, only one parameter (95th percentile) figured among the top ten predictors in terms of differentiating IDH-positive from IDH-negative tumours with relatively lowest area under the curve (AUC) value (Table 5).

Figures 2-4 show examples of parametric maps and 3D VOIs in patients with HGG (Figure 3) and LGG (Figure 4).

Furthermore, decision trees for predicting histological grade and IDH mutation were developed. The sensitivity and specificity for differentiating HGG from LGG were 0.988 and 0.912, respectively (Figure 5). The sensitivity and specificity for differentiating IDH-positive from IDH-negative gliomas were 0.857 and 0.93, respectively (Figure 6).

Figure 5 Result of HGG vs. LGG decision tree analysis. The second line in each box (e.g. 0.78 0.22) is actual proportion of patients with HGG or LGG in each subgroup according to the combination of selected factors; x%, percentage of patients in each subgroup given by combination of factors (sum of percentages in each layer gives 100%). D(5Q), 5th percentile of D; fD*(SD), standard deviation of fD*; K(SD), standard deviation of K; D*(Median), median of D*. HGG, high-grade glioma; LGG, low-grade glioma.

Figure 6 Differentiation of IDH-positive from IDH-negative tumours according to the decision tree analysis. The second line in each box (e.g. 0.78 0.22) is actual proportion of patients with IDH-positive or IDH-negative in each subgroup according to the combination of selected factors; x%, percentage of patients in each subgroup given by combination of factors (sum of percentages in each layer gives 100%). K(75Q), 75th percentile of K; f(Kurt), kurtosis of F; fD*(Median), median of fD*; D(Kurt), kurtosis of D. IDH, isocitrate dehydrogenase; N, negative; P, positive.

Discussion

In this study, we validated the diagnostic yield of multi-b diffusion imaging and ASL techniques, representatives of Gd-contrast-free perfusion methods, in the question of glioma grading and prediction of molecular IDH status using histogram analysis. This revealed significant differences of several histogram parameters between the subgroups.

If we look selectively at the mean values of individual parameters, which are the easiest to interpret, we see results fairly consistent with the literature data and assumptions based on histopathological correlations. Lower mean D values measured in HGG as well as in IDH-negative tumours compared to LGG and IDH-positive tumours, respectively, reflect overall cellularity and nuclear cytoplasmatic ratio (24). This fact is generally in line with previous published studies using quantitative ADC measurements for differentiating LGG from HGG (5,25). Also, the D parameter representing the diffusion coefficient without the non-Gaussian component has been found significantly higher in LGG compared to HGG in studies employing IVIM models for gliomas grading (26). Mean values of perfusion-related parameters D* and fD* were significantly higher in HGG compared to LGG due to higher vascularization of malignant tumours, which is also in line with the review from Li et al. (26). However, Togao et al. did not find D* values to differ significantly between LGG and HGG (9). Conversely, they report larger f values in HGG compared to LGG, which is in contrast with our results not showing significantly different mean f values. This, however, is probably due to differences in the methodological approach, as in that work 2D single-slice region of interest (ROI) analysis was used while placing 2D ROIs on the areas with the highest f-values. 2D ROIs contain significantly fewer voxels than the 3D VOIs of the entire tumour used in our study, and placement of 2D ROIs is substantially influenced by the rater responsible for positioning. This would be consistent with our finding where the maximum f value measured in the whole tumour volume was significantly higher in HGG compared to LGG (q=0.002). The methodology of the aforementioned study also includes additional differences that may explain the discrepancies in the results compared to our study. Data acquisition was performed on a 3T scanner, with variations in the diffusion sequence parameters (TR and TE), and a somewhat different set of measurements was chosen with various b-value settings (13 measurements, 0–1,000 s/mm²), which could independently influence the calculated parameters (27). Lastly, it should be noted that the sample size in that study was smaller than in ours (16 LGG and 29 HGG), which may affect the statistical power of the analyses.

Shen et al. demonstrated a significant decrease in D values and an increase in fD* values in HGG compared to LGG, with the parameter fD* being reported as the strongest predictor of glioma histological grade, evidenced by the highest AUC (11). Our findings are consistent with this, as the 95th percentile of the fD* parameter and the 5th percentile of the D parameter emerge as the two most significant IVIM predictors of glioma grade based on ROC analysis. Additionally, the 5th percentile of D serves as the baseline for differentiating between HGG and LGG within the classification tree. This correlation makes sense considering that in the aforementioned study, the ROI was placed on regions with low D values and high f and D* values, which may approximately correspond to the values identified through histogram analysis. A comprehensive interpretation of the additional histogram parameters is quite challenging due to their large number; however, in most cases, the percentile values exhibit similar behavior to the mean values. Notably, frequent significant differences in the parameters of SD, skewness, and kurtosis are observed, reflecting structural heterogeneity within the tumor tissue (Table S2).

Upon closer examination of the data published to date, we can see that the results are not generally uniform. A meta-analysis by Luo et al. summarizing the results of 6 papers examining 252 tumours identified f as the parameter with the highest pooled specificity and sensitivity (sensitivity 89%, specificity 88%) (12). However, a meta-analysis by Li et al. including 9 papers examining a total of 318 gliomas reported that the f parameter did not differ significantly between LGGs and HGGs in the tumour parenchyma (26). Both these meta-analyses also identify possible reasons for the equivocal and often divergent results. In every case, differences in methodology can be mentioned, as the different numbers of b-values used, as well as their different distributions, have been shown significantly to affect the values of the IVIM parameters obtained (28).

Higher perfusion of HGG has been found in the past by a number of studies using conventional perfusion imaging techniques, such as DSC imaging, where relative cerebral blood volume measurements correlated with the glioma grades and histological findings of increased vascularity of gliomas (29,30). It has been shown that the various IVIM parameters can be related to DSC parameters such as blood volume, mean transit time, and blood flow (8).

The two perfusion methods we investigated, IVIM and ASL, have several advantages over conventional DSC techniques, which could support their wider use in routine practice. Although contrast-enhanced T1 weighted sequences are a routine part of the diagnostic MRI protocol for patients with brain gliomas, the application of contrast agents via injector in two doses for extracellular space presaturation, as required by DSC (31), is not necessary. The described diffusion-based technique is therefore advantageous in terms of simpler technical implementation, reduced costs, and decreased risk of potential complications, such as paravenous contrast administration during high-rate injections. Moreover, contrast perfusion techniques require sufficient temporal resolution, which often comes at the expense of spatial resolution, signal-to-noise ratio, or anatomical coverage. Another advantage of multi-b technique is the ability to quantify additional diffusion parameters, including commonly used ADC values, which would require an additional separate acquisition of diffusion-weighted sequence (8).

In the diagnostic protocol, we included the ASL technique to enable direct comparison of the diagnostic yield of this method with diffusion multi-b imaging within the specific methodology of this study. ASL is a perfusion MRI technique that uses arterial water as an endogenous tracer for perfusion imaging. It has been widely used in studies of glioma grading by quantification of CBF (32,33). However, ASL exhibits a strong dependence on transit time and might underestimate the tumour blood flow with slow blood flowing due to tortuous vessels in the process of angiogenesis (34). Few published papers use both IVIM and ASL techniques for grading gliomas. Shen et al. reported a higher AUC value (0.979) of the fD* parameter for differentiating LGG from HGG compared to CBF quantified by ASL (0.899) (11). Similarly, Lin et al. demonstrated a higher diagnostic yield of IVIM parameters compared to ASL (10). Our findings are consistent with these conclusions. Although we found significant differences between HGG and LGG in several histogram parameters of CBF, none of them is among the ten strongest predictors according to the sums of their sensitivity and specificity to distinguish between those subgroups. Thus, considering also its better spatial resolution, multi-b diffusion technique seems to be a more suitable method for glioma grading.

According to theoretical assumptions, the CBF parameter should generally correlate with the fD* parameter. Although we did not in our analysis observe significant correlation between the mean values of these parameters, we did demonstrate a significant correlation for the 95th percentile of these parameters (i.e., high values of perfusion measured within whole tumours). The SDs of CBF and fD* were also significantly correlated between these parameters. In fact, the SD of fD* proved to be a significant predictor also for subgroup differentiation and was significantly higher in HGG compared to LGG as an expression of higher tissue perfusion heterogeneity.

IDH mutation is the key molecular alteration in the genesis of gliomas (35). It has been discovered that gliomas with IDH mutation are more sensitive to chemoradiation and exhibit better prognosis compared to IDH wild-type tumours (36). Therefore, together with histopathological grading, knowing the status of IDH mutation of gliomas is critical for prognosis and, eventually, personalized treatment of patients (37). In our study, we demonstrated the importance of the investigated techniques for differentiating IDH positive and IDH negative tumours. One of the histograms ASL parameters (95th percentile of CBF) was among the ten most promising predictors, but the parameters derived from multi-b imaging proved to be stronger here as well.

To date, very limited work has been published based upon use of the IVIM diffusivity model to predict the IDH molecular status of gliomas. One such work is that of Yu et al., who demonstrate lower relative D and relative D* values in IDH mutant gliomas in line with our findings, although that study used a different methodological approach based on ROIs and with relativization to contralateral normal brain white matter (38).

Adding high values of b-factors to the diffusion sequence allowed us to calculate the non-tensorial diffusivity kurtosis parameter (K), which is roughly comparable to the mean kurtosis (MK) parameter calculated from multi-shell diffusion technique of DKI (13). Some papers focused on prostate analyses (39), ischemic stroke (40), and brain tumours (41) report strong correlation between MK and K as well as MD and D parameters, respectively. It has been reported that the MK parameter is associated with histological complexity of the examined tissue, with structurally complex tissues showing higher MK values compared to less complex tissues (42). In correlation with this assumption, our results demonstrate significantly increased values of the mean K parameter in HGG and IDH-negative tumours compared to LGG and IDH positive tumours, respectively, and several other histogram K parameters are among the most significant predictors for differentiation of individual subgroups. Specifically, the 95th percentile of the K parameter is the second most powerful predictor of tumour histological grade according to ROC analysis, while the 75th percentile of K is the strongest predictor of the presence of IDH mutation, with this parameter also used at the foundational level of the classification tree. Our results are also in line with a number of studies using DKI imaging for the purpose of glioma grading, and there has been a general consensus across studies that increasing glioma grade correlates with increasing MK values (43,44). Hempel et al., in their study evaluating DKI technique for distinguishing IDH-mutant and IDH-wildtype gliomas, employed an image data analysis method involving 3D segmentation and histogram parameter assessment similar to our study. They found that higher percentile values of the MK parameter (50th, 75th, and 90th) demonstrate slightly better diagnostic performance than the mean MK value (45). An analogous observation can be drawn to our results, where the 75th percentile value of the K parameter shows somewhat higher accuracy (0.813) than the mean K value (0.804) for predicting IDH mutation, although the AUC values are roughly similar in our study (0.837). The significant variations of the K parameter, which approximately correspond to the behaviour of the MK parameter in previous studies, indicate the potential utility of this non-tensorial parameter similarly to conventional multi-shell DKI imaging, whose disadvantage compared to our technique is the more time-consuming data acquisition process. We believe that the integration of isotropic diffusion data over a wide range of b factors offering the possibility of quantifying both IVIM and K parameters into a single non-tensorial multi-b sequence acquired in an acceptable scanning time is desirable for the possibility of this method’s practical application in future.

To identify those areas from which to analyse the multi-b diffusion data, we used semi-automatic segmentation of 3D masks of the tumours including the zone of peritumoral “oedema” surrounding the enhancing tumour core in HGG-appearing tumours, as it is known that this tissue might contain a tumour infiltration (46).

Previous studies utilizing the IVIM method for glioma classification have employed 2D ROI placement in various tumour regions, and it must be noted that the methodological approaches are quite heterogeneous. Some authors favor placing ROIs in areas with low D values or high f or D* values visible on parametric maps (9,11), while in the work of Federau et al. (47), ROIs are positioned in solid tumour regions with the highest signal on DWI (b1000) images or with low ADC values. However, other approaches based solely on structural images for ROI placement have also been described; for example, Wang et al. placed ROIs in the enhancing regions on post-contrast T1 images or in the central part of T2 hyperintense zones in cases of non-enhancing tumors (48). Additionally, some studies utilize histogram analysis of IVIM parameters. However, the histograms were calculated from manually placed 2D ROIs on a single scan across the tumor’s maximum cross-section in T2 images (49) or from ROIs defined on CBF maps derived from ASL data, followed by registration with DWI data (10). We believe that 3D segmentation together with analysis of histograms, as used in our study, could provide more comprehensive and robust results than would analysis of manually placed 2D ROIs (50). To our knowledge, none of the previous studies investigating the application of IVIM for glioma classification employ this methodology; similar approaches can only be found in some publications utilising DKI, e.g., by Hempel et al. (45,51). Techniques based on placing an ROI in a specific area selected a priori (e.g., the highest f value) may overlook important information, such as tumour heterogeneity. This is illustrated by our results of decision trees, which incorporate SD and kurtosis histogram parameters reflecting the heterogeneity of diffusion and contribute significantly to the classification accuracy. A more detailed evaluation of the impact of different methodological approaches to imaging data analysis on the diagnostic yield of multi-b diffusion technique could be a subject of future research.

Due to its rather sizeable time demands, however, the use of 3D segmentation of the entire tumour volume may also represent one of the limitations of the methodology of this work. Nevertheless, we consider it appropriate to use this method at this still relatively early stage of research on this topic in order to verify the behaviour of individual parameters depending on the structure of tumour tissue and to test the overall diagnostic yield. The use of this method in practice could be aided in future by incorporating automated segmentation techniques or fully automated classification algorithms using artificial intelligence methods.

Despite that this study utilizes one of the largest datasets in the field overall, the relatively small group of LGG patients is undoubtedly a limitation. It may limit the power of the statistical analyses and is the reason why we refrained from using a smaller validation set within the classification trees, relying instead on cross-validation to estimate model performance. To enhance the robustness and reliability of our findings, multiple complementary statistical methods were employed, including logistic regression and ROC analysis to assess predictive performance, as well as LASSO regression for variable selection, and classification trees with 10-fold cross-validation to verify classification accuracy. This multimethod approach allowed us to validate results from different perspectives and reduce the risk of model overfitting or instability. In any case, a smaller representation of LGGs is unavoidable in a prospective single-centre study due to their generally lower incidence or postponed surgical treatment. Furthermore, the heterogeneity of the methods used is also a limit of this study from the perspective of interpretation and clinical impact of the results. The calculated parameters may not be entirely reproducible and comparable to those from other studies due to the impacts of differences in the setting of acquisition parameters, especially the number and distribution of b values (27). The differences in data postprocessing, namely application of different curve-fitting models for diffusion data, generally have significant impacts on a method’s capability to differentiate between different tissues and lesions (52). Therefore, it may be challenging to conduct multicentre studies in future, as these would need to deal also with the issue of standardizing methods of data acquisition and analysis to ensure consistent results.

Finally, our study protocol did not encompass conventional DSC perfusion MRI, as we focused on alternative Gd-contrast-free perfusion methods. The direct correlation of this modality with IVIM parameters might further elucidate and support the findings and may be a subject of further studies.

Conclusions

We have shown that both multi-b diffusion imaging and ASL techniques in combination with the analysis of histograms of individual parameters may be useful for differentiating LGG and HGG and for predicting IDH mutation and that multi-b diffusion-derived parameters appear to be more powerful predictors than is CBF calculated from ASL. Further research should focus on the reproducibility issues of individual diffusion data acquisition and analysis techniques. The results presented may also provide a basis for future research using advanced machine learning techniques that could help to implement these techniques in clinical practice.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1182/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1182/dss

Funding: The study was supported by the Ministry of Health of the Czech Republic in cooperation with the Czech Health Research Council (No. NU21-08-00359) to M.D., T.J., Miloš Keřkovský, T.K., Michal Kozubek, E.N. and V.V., and by the Ministry of Health of the Czech Republic Project for Conceptual Development in Research Organizations (University Hospital Brno, Brno, Czech Republic) (No. 65269705) to M.D., T.J., Miloš Keřkovský, T.K., E.N. and V.V.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1182/coif). M.D., T.J., Miloš Keřkovský, T.K., Michal Kozubek, E.N. and V.V. report the funding from the Ministry of Health of the Czech Republic in cooperation with the Czech Health Research Council (No. NU21-08-00359). M.D., T.J., Miloš Keřkovský, T.K., E.N. and V.V. report the funding from the Ministry of Health of the Czech Republic Project for Conceptual Development in Research Organizations (University Hospital Brno, Brno, Czech Republic) (No. 65269705). P.O. reports the employment relationship with the Institute of Biostatistics and Analyses Ltd., Brno, Czech Republic. The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The prospective single-centre study was approved by the Ethics Committee of the University Hospital Brno (No. 21-100620/EK), written informed consent was obtained from all participants. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Scott JN, Brasher PM, Sevick RJ, Rewcastle NB, Forsyth PA. How often are nonenhancing supratentorial gliomas malignant? A population study. Neurology 2002;59:947-9. [Crossref] [PubMed]
2. Satar Z, Hotton G, Samandouras G. Systematic review-Time to malignant transformation in low-grade gliomas: Predicting a catastrophic event with clinical, neuroimaging, and molecular markers. Neurooncol Adv 2021;3:vdab101. [Crossref] [PubMed]
3. van Santwijk L, Kouwenberg V, Meijer F, Smits M, Henssen D. A systematic review and meta-analysis on the differentiation of glioma grade and mutational status by use of perfusion-based magnetic resonance imaging. Insights Imaging 2022;13:102. [Crossref] [PubMed]
4. Ma H, Wang Z, Xu K, Shao Z, Yang C, Xu P, Liu X, Hu C, Lu X, Rong Y. Three-dimensional arterial spin labeling imaging and dynamic susceptibility contrast perfusion-weighted imaging value in diagnosing glioma grade prior to surgery. Exp Ther Med 2017;13:2691-8. [Crossref] [PubMed]
5. Wang QP, Lei DQ, Yuan Y, Xiong NX. Accuracy of ADC derived from DWI for differentiating high-grade from low-grade gliomas: Systematic review and meta-analysis. Medicine (Baltimore) 2020;99:e19254. [Crossref] [PubMed]
6. Duy Hung N, Minh Duc N, Van Anh NT, Thanh Dung L, He DV. Diagnostic performance of diffusion tensor imaging for preo- perative glioma grading. Clin Ter 2021;172:315-21. [Crossref] [PubMed]
7. Xu C, Li C, Xing C, Li J, Jiang X. Efficacy of MR diffusion kurtosis imaging for differentiating low-grade from high-grade glioma before surgery: A systematic review and meta-analysis. Clin Neurol Neurosurg 2022;220:107373. [Crossref] [PubMed]
8. Le Bihan D, Turner R. The capillary network: a link between IVIM and classical perfusion. Magn Reson Med 1992;27:171-8. [Crossref] [PubMed]
9. Togao O, Hiwatashi A, Yamashita K, Kikuchi K, Mizoguchi M, Yoshimoto K, Suzuki SO, Iwaki T, Obara M, Van Cauteren M, Honda H. Differentiation of high-grade and low-grade diffuse gliomas by intravoxel incoherent motion MR imaging. Neuro Oncol 2016;18:132-41. [Crossref] [PubMed]
10. Lin Y, Li J, Zhang Z, Xu Q, Zhou Z, Zhang Z, Zhang Y, Zhang Z. Comparison of Intravoxel Incoherent Motion Diffusion-Weighted MR Imaging and Arterial Spin Labeling MR Imaging in Gliomas. Biomed Res Int 2015;2015:234245. [Crossref] [PubMed]
11. Shen N, Zhao L, Jiang J, Jiang R, Su C, Zhang S, Tang X, Zhu W. Intravoxel incoherent motion diffusion-weighted imaging analysis of diffusion and microperfusion in grading gliomas and comparison with arterial spin labeling for evaluation of tumor perfusion. J Magn Reson Imaging 2016;44:620-32. [Crossref] [PubMed]
12. Luo H, He L, Cheng W, Gao S. The diagnostic value of intravoxel incoherent motion imaging in differentiating high-grade from low-grade gliomas: a systematic review and meta-analysis. Br J Radiol 2021;94:20201321. [Crossref] [PubMed]
13. Kopřivová T, Keřkovský M, Jůza T, Vybíhal V, Rohan T, Kozubek M, Dostál M. Possibilities of Using Multi-b-value Diffusion Magnetic Resonance Imaging for Classification of Brain Lesions. Acad Radiol 2024;31:261-72. [Crossref] [PubMed]
14. Woolrich MW, Jbabdi S, Patenaude B, Chappell M, Makni S, Behrens T, Beckmann C, Jenkinson M, Smith SM. Bayesian analysis of neuroimaging data in FSL. Neuroimage 2009;45:S173-86. [Crossref] [PubMed]
15. Jenkinson M, Smith S. A global optimisation method for robust affine registration of brain images. Med Image Anal 2001;5:143-56. [Crossref] [PubMed]
16. Smith SM. Fast robust automated brain extraction. Hum Brain Mapp 2002;17:143-55. [Crossref] [PubMed]
17. Yushkevich PA, Piven J, Hazlett HC, Smith RG, Ho S, Gee JC, Gerig G. User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage 2006;31:1116-28. [Crossref] [PubMed]
18. Andersson JL, Skare S, Ashburner J. How to correct susceptibility distortions in spin-echo echo-planar images: application to diffusion tensor imaging. Neuroimage 2003;20:870-88. [Crossref] [PubMed]
19. Andersson JLR, Sotiropoulos SN. An integrated approach to correction for off-resonance effects and subject movement in diffusion MR imaging. Neuroimage 2016;125:1063-78. [Crossref] [PubMed]
20. Garyfallidis E, Brett M, Amirbekian B, Rokem A, van der Walt S, Descoteaux M, Nimmo-Smith I. Dipy Contributors. Dipy, a library for the analysis of diffusion MRI data. Front Neuroinform 2014;8:8. [Crossref] [PubMed]
21. Farooq H, Xu J, Nam JW, Keefe DF, Yacoub E, Georgiou T, Lenglet C. Microstructure Imaging of Crossing (MIX) White Matter Fibers from diffusion MRI. Sci Rep 2016;6:38927. [Crossref] [PubMed]
22. Fritzsche KH, Neher PF, Reicht I, van Bruggen T, Goch C, Reisert M, Nolden M, Zelzer S, Meinzer HP, Stieltjes B. MITK diffusion imaging. Methods Inf Med 2012;51:441-8. [Crossref] [PubMed]
23. Wolf I, Vetter M, Wegner I, Nolden M, Bottger T, Hastenteufel M, Schobinger M, Kunert T, Meinzer HP. The medical imaging interaction toolkit (MITK): a toolkit facilitating the creation of interactive software by extending VTK and ITK. In: Medical Imaging 2004: Visualization, Image-Guided Procedures, and Display [Internet]. International Society for Optics and Photonics; 2004:16-27.
24. Yan R, Haopeng P, Xiaoyuan F, Jinsong W, Jiawen Z, Chengjun Y, Tianming Q, Ji X, Mao S, Yueyue D, Yong Z, Jianfeng L, Zhenwei Y. Non-Gaussian diffusion MR imaging of glioma: comparisons of multiple diffusion parameters and correlation with histologic grade and MIB-1 (Ki-67 labeling) index. Neuroradiology 2016;58:121-32. [Crossref] [PubMed]
25. Zhang L, Min Z, Tang M, Chen S, Lei X, Zhang X. The utility of diffusion MRI with quantitative ADC measurements for differentiating high-grade from low-grade cerebral gliomas: Evidence from a meta-analysis. J Neurol Sci 2017;373:9-15. [Crossref] [PubMed]
26. Li WF, Niu C, Shakir TM, Chen T, Zhang M, Wang Z. An evidence-based approach to assess the accuracy of intravoxel incoherent motion imaging for the grading of brain tumors. Medicine (Baltimore) 2018;97:e13217. [Crossref] [PubMed]
27. Conklin J, Heyn C, Roux M, Cerny M, Wintermark M, Federau C. A Simplified Model for Intravoxel Incoherent Motion Perfusion Imaging of the Brain. AJNR Am J Neuroradiol 2016;37:2251-7. [Crossref] [PubMed]
28. Chabert S, Verdu J, Huerta G, Montalba C, Cox P, Riveros R, Uribe S, Salas R, Veloz A. Impact of b-Value Sampling Scheme on Brain IVIM Parameter Estimation in Healthy Subjects. Magn Reson Med Sci 2020;19:216-26. [Crossref] [PubMed]
29. Hilario A, Ramos A, Perez-Nuñez A, Salvador E, Millan JM, Lagares A, Sepulveda JM, Gonzalez-Leon P, Hernandez-Lain A, Ricoy JR. The added value of apparent diffusion coefficient to cerebral blood volume in the preoperative grading of diffuse gliomas. AJNR Am J Neuroradiol 2012;33:701-7. [Crossref] [PubMed]
30. Hakyemez B, Erdogan C, Ercan I, Ergin N, Uysal S, Atahan S. High-grade and low-grade gliomas: differentiation by using perfusion MR imaging. Clin Radiol 2005;60:493-502. [Crossref] [PubMed]
31. Hu LS, Baxter LC, Pinnaduwage DS, Paine TL, Karis JP, Feuerstein BG, Schmainda KM, Dueck AC, Debbins J, Smith KA, Nakaji P, Eschbacher JM, Coons SW, Heiserman JE. Optimized preload leakage-correction methods to improve the diagnostic accuracy of dynamic susceptibility-weighted contrast-enhanced perfusion MR imaging in posttreatment gliomas. AJNR Am J Neuroradiol 2010;31:40-8. [Crossref] [PubMed]
32. Kim HS, Kim SY. A prospective study on the added value of pulsed arterial spin-labeling and apparent diffusion coefficients in the grading of gliomas. AJNR Am J Neuroradiol 2007;28:1693-9. [Crossref] [PubMed]
33. Warmuth C, Gunther M, Zimmer C. Quantification of blood flow in brain tumors: comparison of arterial spin labeling and dynamic susceptibility-weighted contrast-enhanced MR imaging. Radiology 2003;228:523-32. [Crossref] [PubMed]
34. Petersen ET, Zimine I, Ho YC, Golay X. Non-invasive measurement of perfusion: a critical review of arterial spin labelling techniques. Br J Radiol 2006;79:688-701. [Crossref] [PubMed]
35. Molinaro AM, Taylor JW, Wiencke JK, Wrensch MR. Genetic and molecular epidemiology of adult diffuse glioma. Nat Rev Neurol 2019;15:405-17. [Crossref] [PubMed]
36. Pirozzi CJ, Yan H. The implications of IDH mutations for cancer development and therapy. Nat Rev Clin Oncol 2021;18:645-61. [Crossref] [PubMed]
37. Guo D, Jiang B. Noninvasively evaluating the grade and IDH mutation status of gliomas by using mono-exponential, bi-exponential diffusion-weighted imaging and three-dimensional pseudo-continuous arterial spin labeling. Eur J Radiol 2023;160:110721. [Crossref] [PubMed]
38. Yu M, Ge Y, Wang Z, Zhang Y, Hou X, Chen H, Chen X, Ji N, Li X, Shen H. The diagnostic efficiency of integration of 2HG MRS and IVIM versus individual parameters for predicting IDH mutation status in gliomas in clinical scenarios: A retrospective study. J Neurooncol 2024;167:305-13. [Crossref] [PubMed]
39. Malagi AV, Netaji A, Kumar V, Baidya Kayal E, Khare K, Das CJ, Calamante F, Mehndiratta A. IVIM-DKI for differentiation between prostate cancer and benign prostatic hyperplasia: comparison of 1.5 T vs. 3 T MRI. MAGMA 2022;35:609-20. [Crossref] [PubMed]
40. Pavilla A, Gambarota G, Signaté A, Arrigo A, Saint-Jalmes H, Mejdoubi M. Intravoxel incoherent motion and diffusion kurtosis imaging at 3T MRI: Application to ischemic stroke. Magn Reson Imaging 2023;99:73-80. [Crossref] [PubMed]
41. Loução R, Oros-Peusquens AM, Langen KJ, Ferreira HA, Shah NJ. A Fast Protocol for Multiparametric Characterisation of Diffusion in the Brain and Brain Tumours. Front Oncol 2021;11:554205. [Crossref] [PubMed]
42. Pang H, Ren Y, Dang X, Feng X, Yao Z, Wu J, Yao C, Di N, Ghinda DC, Zhang Y. Diffusional kurtosis imaging for differentiating between high-grade glioma and primary central nervous system lymphoma. J Magn Reson Imaging 2016;44:30-40. [Crossref] [PubMed]
43. Zhang J, Chen X, Chen D, Wang Z, Li S, Zhu W. Grading and proliferation assessment of diffuse astrocytic tumors with monoexponential, biexponential, and stretched-exponential diffusion-weighted imaging and diffusion kurtosis imaging. Eur J Radiol 2018;109:188-95. [Crossref] [PubMed]
44. Van Cauter S, Veraart J, Sijbers J, Peeters RR, Himmelreich U, De Keyzer F, Van Gool SW, Van Calenbergh F, De Vleeschouwer S, Van Hecke W, Sunaert S. Gliomas: diffusion kurtosis MR imaging in grading. Radiology 2012;263:492-501. [Crossref] [PubMed]
45. Hempel JM, Schittenhelm J, Brendle C, Bender B, Bier G, Skardelly M, Tabatabai G, Castaneda Vega S, Ernemann U, Klose U. Histogram analysis of diffusion kurtosis imaging estimates for in vivo assessment of 2016 WHO glioma grades: A cross-sectional observational study. Eur J Radiol 2017;95:202-11. [Crossref] [PubMed]
46. Villanueva-Meyer JE, Mabray MC, Cha S. Current Clinical Brain Tumor Imaging. Neurosurgery 2017;81:397-415. [Crossref] [PubMed]
47. Federau C, Meuli R, O'Brien K, Maeder P, Hagmann P. Perfusion measurement in brain gliomas with intravoxel incoherent motion MRI. AJNR Am J Neuroradiol 2014;35:256-62. [Crossref] [PubMed]
48. Wang X, Chen XZ, Shi L, Dai JP. Glioma grading and IDH1 mutational status: assessment by intravoxel incoherent motion MRI. Clin Radiol 2019;74:651.e7-651.e14. [Crossref] [PubMed]
49. Kusunoki M, Kikuchi K, Togao O, Yamashita K, Momosaka D, Kikuchi Y, Kuga D, Hata N, Mizoguchi M, Iihara K, Suzuki SO, Iwaki T, Akamine Y, Hiwatashi A. Differentiation of high-grade from low-grade diffuse gliomas using diffusion-weighted imaging: a comparative study of mono-, bi-, and stretched-exponential diffusion models. Neuroradiology 2020;62:815-23. [Crossref] [PubMed]
50. Mesropyan N, Mürtz P, Sprinkart AM, Block W, Luetkens JA, Attenberger U, Pieper CC. Comparison of different ROI analysis methods for liver lesion characterization with simplified intravoxel incoherent motion (IVIM). Sci Rep 2021;11:22752. [Crossref] [PubMed]
51. Hempel JM, Brendle C, Bender B, Bier G, Kraus MS, Skardelly M, Richter H, Eckert F, Schittenhelm J, Ernemann U, Klose U. Diffusion kurtosis imaging histogram parameter metrics predicting survival in integrated molecular subtypes of diffuse glioma: An observational cohort study. Eur J Radiol 2019;112:144-52. [Crossref] [PubMed]
52. Keil VC, Mädler B, Gielen GH, Pintea B, Hiththetiya K, Gaspranova AR, Gieseke J, Simon M, Schild HH, Hadizadeh DR. Intravoxel incoherent motion MRI in the brain: Impact of the fitting model on perfusion fraction and lesion differentiability. J Magn Reson Imaging 2017;46:1187-99. [Crossref] [PubMed]