Impact of Free Water Correction on the Mean Diffusivity of Glioblastoma is Regional and b-value Dependent: A Retrospective Study on Open-source Datasets

Abstract

Purpose

To investigate the effects of free-water (FW) correction on mean diffusivity (MD) values in glioblastomas using diffusion tensor imaging (DTI) with different b-value settings. We also examined regional differences between contrast-enhanced (CE) tumor and peritumoral fluid-attenuated inversion recovery (FLAIR) hyperintense regions.

Methods

DTI data from 787 participants with glioblastoma were obtained from 2 open-source datasets: the University of California, San Francisco Preoperative Diffuse Glioma MRI (UCSF-PDGM; b-value, 2000) and the Multi-parametric Magnetic Resonance Imaging Scans for De Novo Glioblastoma Patients from the University of Pennsylvania Health System (UPENN-GBM; b-value, 1000). Conventional and FW-corrected MD maps were generated, and median MD values were calculated from CE and FLAIR hyperintense regions. The impact of FW correction was assessed via the Wilcoxon matched-pairs signed rank test, Bland–Altman plot, and Spearman’s correlation analysis. Multivariate Cox regression was used to evaluate prognostic significance, adjusting for age and CE volume.

Results

Both median MD values were significantly lower in the CE regions than in the FLAIR hyperintense regions. In the UCSF-PDGM dataset, FW correction caused a greater reduction of MD in the CE regions, whereas the greater reduction occurred in the FLAIR hyperintense regions in the UPENN-GBM dataset. Bland–Altman plot also indicated lesions with lower MD values showing greater FW correction in the UCSF-PDGM dataset, whereas those with relatively high MD values showed greater FW correction in the UPENN-GBM dataset. The correlation between both MD values was high (Rs > 0.98). Neither conventional MD nor FW-corrected MD values in the CE regions were significantly associated with overall survival. (Ps > 0.30).

Conclusion

A paradoxical relationship of MD values with and without FW correction was identified depending on regional characteristics and b-value settings, underscoring the importance of acquisition parameters when interpreting FW correction.

Introduction

Glioblastoma (GBM) is the most aggressive primary brain tumor, with a median survival of only 15 months despite intensive treatment that usually comprises surgery and subsequent radiochemotherapy.¹ Radiologically, the margins of the post-contrast enhanced (CE) regions on MRI do not represent true tumor borders.² Most tumors recur in the peritumoral areas even after the CE regions have been completely resected.³ These areas often showed hyperintensity on fluid-attenuated inversion recovery (FLAIR) images; nevertheless, distinguishing tumor infiltration from peritumoral edema is clinically impossible. Nonetheless, accurately recognizing intratumoral features is necessary to predict tumor activity and determine the optimal surgical margin, as well as plan appropriate radiation therapy.

The apparent diffusion coefficient (ADC) calculated from diffusion weighted imaging (DWI) reflects water diffusion within tissues and correlates with cell density and proliferation activity.⁴ The ADC values, especially within the post-operatively residual lesions, were suggested to be associated with the prognosis for patients with GBM; however, these values may not correctly reflect the activity of tumors because of the edematous effects and tumor infiltration. It was suggested in a previous study that an ADC histogram analysis of the tumor lesions could be used to stratify tumor and edematous effects using 2 normal distribution curves (a double Gaussian model).⁵ The value of the lower peak of the histogram reflects cell density by mitigating the influence of the extracellular fluid component and could be used to predict outcomes more accurately compared with conventional ADC values.^5,6 ADC histogram analysis may have enabled the separation of the tumor infiltration and edematous or necrotic tissue; nevertheless, it may results in failure when the tumor lesion is quite small because of its mathematical nature.

Diffusion tensor imaging (DTI) is an advanced MR technique that can be used to detect the Brownian motion of water molecules in equal or more than 6 directions.⁷ Mean diffusivity (MD), which shows similar behavior to ADC, derived from DTI, is also often used as a metric for GBM evaluation. Recently, the elimination of free water (FW) has become measurable from DTI data, and FW-corrected MD value may reflect tissue characteristics after suppressing the extracellular free-water component.⁸ FW-corrected MD values may help mitigate peritumoral edema and highlight tumor infiltration; nonetheless, brain tumors were evaluated using this approach in only a few studies.^9–11

Considering that FW-corrected MD values are a crucial metric in future studies for predicting actual tumor activity and treatment planning, clarifying the relationship between MD values with and without FW correction is necessary. This relationship may further depend on the evaluated regions of tumors, as the elimination effect may vary across different tumor areas. Furthermore, water diffusivity is influenced by b-value settings; however, there are no clear reports regarding the degree of change between the MD values with and without FW correction in different b-values. Therefore, the purpose of this study was to investigate the FW-corrected MD values of GBMs with different b-value settings using large open-source datasets.

Materials and Methods

Written consent from the participants and ethics board review or approval was not required since the dataset used in this study was publicly available. This study was conducted following the principles of the Declaration of Helsinki and was structured in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines.¹²

Participants

MRI data of GBMs were retrieved from 2 publicly available datasets in The Cancer Imaging Archive (https://www.cancerimagingarchive.net/): the University of California, San Francisco Preoperative Diffuse Glioma MRI (UCSF-PDGM) and the multi-parametric MRI scans for de novo GBM patients from the University of Pennsylvania Health System (UPENN-GBM).^13–15 Participants diagnosed with isocitrate dehydrogenase wild-type glioma were selected from those whose isocitrate dehydrogenase status had been confirmed, even if they were previously classified as having grade 2 or 3 diffuse astrocytomas under the 4th edition of the World Health Organization Classification of the Tumours of the Central Nervous System. Among them, participants whose DTI data were available were included, while those whose subsequent FW correction was not successful were excluded. For survival analysis, only participants with documented survival outcomes indicating either death or censored status at the study endpoint were included. The selection process is illustrated in Fig. 1.

Fig. 1.

Data collection flowchart.DTI, diffusion tensor imaging; FW, free water; IDH, isocitrate dehydrogenase; UCSF-PDGM, University of California, San Francisco Preoperative Diffuse Glioma MRI; UPENN-GBM, glioblastoma patients from the University of Pennsylvania Health System.

MRI and demographic data extraction

In this study, preoperative CE–T1 weighted imaging (T1WI), T2-FLAIR imaging, DTI, and tumor segmentation data were extracted. Tumor segmentation was already performed and registered to CE-T1WI when the data was retrieved from the Cancer Imaging Archive. Tumor segmentation included CE ROIs, including CE areas but not necrotic areas, and FLAIR ROIs, including T2-FLAIR hyperintense areas but not overlapping with CE ROIs. The volume of these ROIs was calculated. All DTI data of UCSF-PDGM were obtained with 55 diffusion-encoding directions at TR = 8400, TE = 73, and b = 2000 s/mm² and one b = 0 s/mm² image from a 3T machine, Discovery 750 (GE Healthcare, Chicago, IL, USA). DTI data of UPENN-GBM was obtained with various diffusion-encoding directions (12–30) at TR = 3100–7600, TE = 83–121, and b = 1000 s/mm² and b = 0 s/mm² images from a 1.5T (11%) or a 3T machine (89%). The TrioTim (Siemens Healthineers, Erlangen, Germany) of 3T was the most frequently used machine (79%). Detailed DTI parameters are shown on the Cancer Imaging Archive website. Participants’ demographic data, including age, sex, overall survival, and final status (dead or censored), were also retrieved.

DTI preprocessing and free-water correction

All DTI data were preprocessed using the MRtrix3 software (https://www.mrtrix.org) to minimize distortions and artifacts before the FW correction. The procedure involved denoising, removal of Gibbs ringing artifacts, and correction for motion, eddy currents, and bias fields.¹⁶ Subsequently, MD maps were created from the preprocessed DTI data. FW-correction was also performed using the preprocessed DTI data to create FW-corrected MD maps.^8,17 The MD and FW-corrected MD maps were coregistered and registered to the CE-T1WI with 6 degrees of freedom rigid transformation. The MD map was created, and images were registered using the FMRIB Software Library version 6.0 (FSL; Oxford, UK; www.fmrib.ox.ac.uk/fsl). Median MD values and median FW-corrected MD values within each tumor segmentation (CE ROIs and FLAIR ROIs) were calculated. Median values were selected for robustness against outliers, such as those arising from intratumoral necrosis, cysts, or numerical instabilities in the FW-correction process. An example of MD and FW-corrected MD maps with segmentation data is shown in Fig. 2. MD values were extracted from the CE and FLAIR ROIs. A double Gaussian mixture model was applied to the MD histograms using a nonlinear least-squares regression algorithm in MATLAB (Version R2024a; MathWorks, Natick, MA, USA).^5,18 The lower peak value was defined as the mean of the Gaussian component with the lower MD value. This peak-based approach was chosen because it is more robust to variations in ROI size and partial volume effects at the margins compared to simple mean or median values.

Fig. 2.

An example of the MD and FW-corrected MD maps and segmentation data of GBM. (a), (b), (c), and (d) indicate MD map, FW-corrected MD map, CE-T1WI, and CE-T1WI with CE (red) and FLAIR hyperintense regions (blue), respectively. All images were co-registered.CE, contrast enhanced; FLAIR, fluid-attenuated inversion recovery; FW, free water; GBM, glioblastoma; MD, mean diffusivity; T1WI, T1 weighted imaging.

Statistical analysis

Non-parametric analysis methods were chosen because the Shapiro–Wilk test confirmed that the MD values within the ROIs did not follow a normal distribution in many cases. Consequently, the median was employed as a more stable representative value than the mean, as it is less sensitive to potential mathematical outliers generated during FW processing. The median MD values with or without FW-correction within CE and FLAIR ROIs and the subtracted values between the MD and FW-corrected MD values were compared using the Wilcoxon matched-pairs signed rank test. Spearman’s rank correlation coefficient between the median MD and FW-corrected MD was calculated. The Bland–Altman plot between the median of both MD values was evaluated using CE and FLAIR ROIs. The same analyses were repeated using the lower peak values of MD derived from the double Gaussian model. Additionally, a subgroup analysis was conducted within the UPENN-GBM dataset by including only participants scanned with the TrioTim 3T system, which was the most frequently used scanner. Multivariate Cox regression of the median MD or FW-corrected MD value was performed to calculate the hazard ratio using the age and CE ROI volume as covariates.

In addition to the MD metrics, the FW fraction (f) was extracted for each ROI. To evaluate the FW correction on the diffusion metrics, the relationship between the FW fraction and the MD metrics (original MD, FW-corrected MD, and the subtraction [FW-corrected MD − original MD]) across the UCSF-PDGM (b = 2000 s/mm²) and UPENN-GBM (b = 1000 s/mm²) datasets were analyzed.

Two-tailed P values < 0.05 denoted statistical significance. All statistical analyses were performed using GraphPad Prism version 10.4.1 software (GraphPad Software, San Diego, CA, USA).

Results

Among the 897 participants (UCSF-PDGM, 398; UPENN-GBM, 499) with confirmed isocitrate dehydrogenase wild-type glioma, 816 participants (UCSF-PDGM, 392; UPENN-GBM, 424) who underwent DTI acquisition were selected, while 29 participants (UCSF-PDGM, 1; UPENN-GBM, 28) whose FW-correction was not successful due to excessive processing time or severe image artifacts were excluded. Thus, 787 participants (UCSF-PDGM, 391; UPENN-GBM, 396) were included in the final FW-correction analysis. However, 22 participants (UCSF-PDGM, 18; UPENN-GBM, 4) did not have CE ROI because no CE regions were identified on CE-T1WI. Nevertheless, 786 participants (UCSF-PDGM, 390; UPENN-GBM, 396) with documented overall survival outcomes were included in the survival analysis. Table 1 shows the demographic data of the included participants.

Table 1.

Demographic data of participants

	Total (n = 787)	UCSF-PDGM (n = 391)	UPENN-GBM (n = 396)
Gender (M/F)	471/316	232/159	239/157
Age (year, mean ± SD)	62 ± 11.8	62 ± 12.0	63 ± 11.5
Median Survival (days)	454	512	412

F, female; M, male; SD; standard deviation; UCSF-PDGM, University of California, San Francisco Preoperative Diffuse Glioma MRI; UPENN-GBM, glioblastoma patients from the University of Pennsylvania Health System.

Table 2 shows the median MD and FW-corrected MD values, as well as the subtraction of MD values within the CE and FLAIR ROIs. The median of both values was significantly lower in the CE ROIs than in the FLAIR ROIs in both UCSF-PDGM and UPENN-GBM datasets. Subtraction between MD and FW-corrected MD was significantly greater in the CE ROIs than in the FLAIR ROIs in the UCSF-PDGM dataset, whereas it was significantly greater in the FLAIR ROIs than in the CE ROIs in the UPENN-GBM dataset.

Table 2.

MD, FW-corrected MD, and subtraction of MD in different ROIs

		CE ROI	FLAIR ROI	P value
UCSF-PDGM	MD (×10−4 mm2/s)	8.83 (7.67–10.1)	10.4 (9.07–11.7)	<0.001*
	FW-corrected MD (×10−4 mm2/s)	7.06 (5.60–8.85)	9.18 (7.34–11.0)	<0.001*
	Subtraction of MD (×10−4 mm2/s)	1.71 (1.19–1.98)	1.23 (0.77–1.85)	<0.001*
UPENN-GBM	MD (×10−4 mm2/s)	10.7 (9.42–12.1)	11.9 (10.4–13.6)	<0.001*
	FW-corrected MD (×10−4 mm2/s)	7.10 (5.92–8.55)	7.92 (6.40–9.83)	<0.001*
	Subtraction of MD (×10−4 mm2/s)	3.46 (3.19–3.67)	3.71 (3.46–4.01)	<0.001*

Medians, with interquartile range.

^*Statistically significant. CE, post-contrast enhanced; FLAIR, fluid attenuated inversion recovery; FW, free water; MD, mean diffusivity; UCSF-PDGM, University of California, San Francisco Preoperative Diffuse Glioma MRI; UPENN-GBM, glioblastoma patients from the University of Pennsylvania Health System.

The correlation between MD and FW-corrected MD values is shown in Fig. 3, while Fig. 4 shows the Bland–Altman plot. The correlation coefficients (Rs) between median MD and FW-corrected MD values were 0.99 and 0.99 in the CE and FLAIR ROIs, respectively, in the UCSF-PDGM dataset, while the values were 0.98 and 0.98 in the CE and FLAIR ROIs, respectively, in the UPENN-GBM dataset.

Fig. 3.

Scatter plot of MD and FW-corrected MD in (a) UCSF-PDGM and (b) UPENN-GBM datasets. In the CE and FLAIR ROIs, a significant positive correlation between MD and FW-corrected MD values is shown in both (a) UCSF-PDGM and (b) UPENN-GBM datasets.CE, contrast enhanced; FLAIR, fluid-attenuated inversion recovery; FW, free water; GBM, glioblastoma; MD, mean diffusivity; UCSF-PDGM, University of California, San Francisco Preoperative Diffuse Glioma MRI; UPENN-GBM, glioblastoma patients from the University of Pennsylvania Health System.

Fig. 4.

Bland-Altman plot showing the comparison of the MD and FW-corrected MD. The difference between MD and FW-corrected MD becomes wider as MD declines in both (a) UCSF-PDGM and (b) UPENN-GBM datasets. This tendency is equally identified in CE and FLAIR ROIs in (a) the UCSF-PDGM dataset; however, wider declines were identified in the FLAIR ROIs than in the CE ROIs in (b) the UPENN-GBM dataset. The 2 lines in both figures indicate the 95% limit of agreement of CE ROI data.CE, contrast enhanced; FLAIR, fluid-attenuated inversion recovery; FW, free water; GBM, glioblastoma; MD, mean diffusivity; UCSF-PDGM, University of California, San Francisco Preoperative Diffuse Glioma MRI; UPENN-GBM, glioblastoma patients from the University of Pennsylvania Health System.

In the Bland–Altman plot, lesions with lower MD values, which were generally located in the CE ROIs, showed greater FW correction in the UCSF-PDGM dataset with a b-value of 2000. In contrast, lesions with relatively high MD values, which were generally located in the FLAIR ROIs, showed greater FW correction in the UPENN-GBM dataset with a b-value of 1000. The 95% limits of agreement for the difference between MD and FW-corrected MD were −0.27 × 10⁻³ to −0.043 × 10⁻³ mm²/s in the UCSF-PDGM dataset and −0.46 × 10⁻³ to −0.22 × 10⁻³ mm²/s in the UPENN-GBM dataset.

In the analyses using the lower peak MD values, the values were significantly lower in the CE ROIs than in the FLAIR ROIs in both UCSF-PDGM and UPENN-GBM datasets (Supplementary Table 1). The correlation between MD and lower peak of MD values is shown in Supplementary Fig. 1, while Supplementary Fig. 2 shows the Bland–Altman plot. In the subgroup analysis of the UPENN-GBM dataset restricted to the TrioTim scanner (n = 313), the findings remained consistent with the overall results (Supplementary Table 2). The correlation between MD and FW-corrected MD values and the Bland–Altman plot obtained from TrioTim are shown in Supplementary Fig. 3.

The analysis of the relationship between the FW fraction and MD metrics revealed that both original MD and FW-corrected MD increased as the FW fraction increased in both datasets (Fig. 5). In contrast, the behavior of the MD difference (subtraction) following FW correction varied considerably depending on the b-value. In the UCSF-PDGM dataset (b = 2000), the subtraction showed an upward trend (approaching zero) as the FW fraction increased, indicating that the magnitude of the correction effect decreased in regions with higher FW. Conversely, in the UPENN-GBM dataset (b = 1000), the subtraction remained substantially constant across the FW spectrum, maintaining a large correction effect even at high FW fractions.

Fig. 5.

Relationship between the FW fraction and MD metrics across different b-values. Scatter plots demonstrate the relationship between the FW fraction (x-axis) and MD metrics (y-axis). The upper part of each plot (positive y-axis) shows the FW-corrected and original MD, while the lower part (negative y-axis) shows the subtraction (FW-corrected MD – original MD). (a) and (b) represent the CE ROI (red) and FLAIR hyperintense ROI (blue) in the UCSF-PDGM dataset (b = 2000 s/mm²), respectively. (c) and (d) represent the CE ROI (red) and FLAIR hyperintense ROI (blue) in the UPENN-GBM dataset (b = 1000 s/mm²), respectively. In all panels, MD values increase with higher FW content. However, the behavior of the subtraction varies significantly: in the UCSF-PDGM dataset (a, b), the subtraction magnitude decreases (approaches zero) as the FW fraction increases, whereas in the UPENN-GBM dataset (c, d), it remains relatively constant across the FW spectrum.CE, contrast enhanced; FLAIR, fluid-attenuated inversion recovery; FW, free water; GBM, glioblastoma; MD, mean diffusivity; ROI, region of interest; UCSF-PDGM, University of California, San Francisco preoperative diffuse glioma MRI; UPENN-GBM, glioblastoma patients from the University of Pennsylvania Health System.

In either dataset, median MD or FW-corrected MD values within the CE ROIs were not significantly associated with prognosis when controlling for age and CE ROI volume as covariates (Ps > 0.30). Detailed results of the survival analysis are provided in Supplementary Table 3.

Discussion

In this study, we investigated the FW correction of MD values in the tumor and peritumoral regions of GBMs with different b-value settings. We showed that MD and FW-corrected MD values, along with lower peak MD values, were significantly lower in CE ROIs than in FLAIR ROIs, even when using DTI data with b-values of either 1000 or 2000. In contrast, when using DTI with a b-value of 2000, FW-correction caused a greater reduction of MD value in CE ROIs than in the FLAIR ROIs, whereas using DTI with a b-value of 1000, a greater reduction of MD values was observed in FLAIR ROIs than in the CE ROIs. This phenomenon was further supported by the interaction between the FW fraction and correction magnitude, which exhibited a paradoxical upward trend at b = 2000. Finally, neither median MD nor FW-corrected MD showed a significant association with overall survival in our cohort.

In this study, the degree of MD reduction after FW correction markedly differed depending on whether the diffusion data were acquired at b =  2000 s/mm² (UCSF-PDGM) or b =  1000 s/mm² (UPENN-GBM). Specifically, in the UCSF-PDGM dataset, tumor lesions in the CE regions exhibited a greater reduction of MD values following FW correction than those in the FLAIR hyperintense regions. In contrast, an opposite pattern was found in the UPENN-GBM dataset, with greater MD reduction in the FLAIR hyperintense regions than in the CE regions. One potential explanation for this discrepancy was that under higher b-value conditions, such as b = 2000, the diffusion-weighted signal became more sensitive to restricted diffusion. In GBM, the CE region typically exhibited high cellular density, which caused greater signal attenuation owing to restricted diffusion. Therefore, part of the FW component in the CE region was already suppressed during image acquisition. When FW correction was subsequently applied, the relative decrease in MD appeared quite large, which substantially highlighted areas of truly restricted diffusion.^19,20 Conversely, at lower b-values, such as b = 1000, much of the FW signal in edematous or fluid-rich tissue, including the FLAIR hyperintense regions, remained prominent before FW correction. Because this FW component had not been substantially attenuated at acquisition, the subsequent FW correction remarkably reduced the MD values. Thus, in these FLAIR-dominant areas, a more pronounced post-correction change in MD values may be shown, indicating that a large FW signal was effectively removed.

To further corroborate this b-value-dependent behavior, the analysis of the FW fraction (Fig. 5) highlights the importance of independently interpreting both the FW fraction (f) and FW-corrected MD.⁸ While the FW fraction is an intuitive metric reflecting the estimated degree of edema, our analysis demonstrates that the magnitude of mathematical correction is not a simple linear function of the FW fraction, but is highly b-value dependent.²¹ Intuitively, higher FW content should necessitate a larger mathematical correction. However, a paradoxical upward trend was observed in the UCSF-PDGM dataset (b = 2000), where the correction magnitude decreased as the FW fraction increased. As discussed above, this paradox may be explained by signal attenuation physics: at b = 2000, the highly mobile FW signal may be substantially attenuated during acquisition. Consequently, the original MD is already biased toward the restricted tissue compartment, leaving the algorithm with little residual FW signal to subtract in high-FW regions. In contrast, at b = 1000 (UPENN-GBM dataset), substantial FW signal persists, which may result in a consistently large correction across the FW spectrum.

While a uniform imaging protocol is employed in the UCSF-PDGM dataset, the UPENN-GBM data were acquired using multiple scanners and parameter sets, including variations in magnetic field strength (1.5T and 3T) and diffusion-encoding directions (12–30 directions). These variations, particularly the inclusion of 1.5T data and the use of fewer encoding directions compared to the UCSF-PDGM dataset (55 directions), could potentially introduce biases in MD estimation and the robustness of FW correction. Nonetheless, subgroup analysis restricted to the most frequently used scanner in the UPENN-GBM dataset (TrioTim) showed findings consistent with the overall results. This suggests that the observed paradoxical trend is primarily driven by b-value–dependent signal attenuation characteristics rather than differences in magnetic field strength, scanner hardware, or the number of encoding directions. Additionally, while MD and FW-corrected MD were highly correlated, the amount of reduction was not uniform across regions. This indicates that FW correction reflects regional tissue characteristics that are not fully captured by conventional MD. Therefore, future multicenter studies combining data with different b-values require careful attention, and advanced harmonization techniques, such as ComBat, are recommended to reduce these protocol-dependent biases.²²

Regarding the histogram-based analysis, both lower peak MD and FW-corrected MD share a common goal: isolating the hypercellular tumor component by mitigating the confounding effects of extracellular edema. Our results showed that lower peak MD was more consistent across datasets and b-value settings than FW-corrected MD. This suggests that while voxel-wise FW correction in single-shell data is mathematically unstable,²³ the statistical extraction of the lower peak offers a more numerically robust metric. However, a potential limitation of lower peak MD is its dependence on ROI size. The double Gaussian model requires a sufficient number of voxels for stable fitting; thus, its reliability may be compromised in very small ROIs where the histogram distribution is poorly defined.²⁴

Both MD and FW-corrected MD values were lower in the CE ROIs than in the FLAIR ROIs, possibly reflecting the higher malignancy of the CE regions. Furthermore, FW correction may reduce the influence of extracellular water content in a manner similar to high b-value DTI, which was used to distinguish low-grade from high-grade gliomas.^25,26 Given that CE ROIs were expected to represent more aggressive tumor regions, the MD and FW-corrected MD values in these areas were consistent with their higher cellularity and restricted diffusion.

In this study, the MD or FW-corrected MD value was not significantly associated with prognosis, possibly because it was necessary to use preoperative images from open-source datasets. The extension of the tumor resection would have more influenced the prognosis than the preoperative MD or FW-corrected MD measurements. In some previous studies, MD values were also used in recurrent lesions to predict prognosis, suggesting their potential utility in post-treatment assessment.^5,6,27 Therefore, further investigation is required to evaluate the contribution of both conventional MD and FW-corrected MD to prognostic prediction using postoperative data.

There were some limitations in this study that should be addressed. First, a retrospective analysis based on open-source data was employed. Notably, while the UCSF-PDGM dataset had standardized imaging parameters, the MRI acquisition parameters, and scanners in the UPENN-GBM dataset were not uniform. Further prospective validation with an identical protocol is required. However, because approximately 800 participants with GBM were analyzed in this study, the results would have indicated that the reduction in MD with FW correction differs depending on the measurement location within the tumor and on the b-value settings. Second, neither conventional nor FW-corrected MD values showed significant prognostic value, which may be due to the lack of critical clinical covariates. Important factors such as the extent of resection, O⁶-methylguanine-DNA methyltransferase promoter status, and postoperative treatment details were not consistently available across the open-source datasets and could not be included in our survival analysis. Third, single-shell FW correction is mathematically ill-posed and prone to bias.^23,28 Furthermore, high b-values are affected by non-Gaussian kurtosis effects. Although we lacked multi-shell data for comparison, our large sample size enabled us to identify robust, systematic trends in these biases. Fourth, using the findings alone, we could not conclude whether the FW-corrected MD value would be a superior metric compared with the conventional MD value. Further validation is required, particularly in assessing MD and FW-corrected MD values using postoperative images.

Conclusion

In this study, the paradoxical relationship of MD values with and without FW correction in different GBM regions and b-value settings was identified. Potential FW components may have influenced the degree of the MD reduction in GBMs. These findings would support future research where FW correction is used to evaluate the tumor cell density or activity and to determine the optimal surgical margin, as well as to plan appropriate radiation therapy.

Supplementary Materials

Supplementary figures are available online.

Supplementary Tables

Supplementary Table 1.

Lower peaks of MD in different ROIs

Supplementary Table 2.

MD, FW-corrected MD, and lower peaks of MD obtained from TrioTim in the UPENN-GBM dataset

Supplementary Table 3.

Multivariate Cox regression controlling for age and CE volume

Supplementary Figures

Supplementary Figure 1.

Scatter plot of MD and lower peak of MD in (a) UCSF-PDGM and (b) UPENN-GBM datasets

CE, contrast enhanced; FLAIR, fluid-attenuated inversion recovery; MD, mean diffusivity; UCSF-PDGM, University of California, San Francisco Preoperative Diffuse Glioma MRI; UPENN-GBM, glioblastoma patients from the University of Pennsylvania Health System.

Supplementary Figure 2.

Bland-Altman plot showing the comparison of the MD and lower peaks of MD in (a) UCSF-PDGM and (b) UPENN-GBM datasets. The two lines in both figures indicate the 95% limit of agreement of CE ROI data.

CE, contrast enhanced; FLAIR, fluid-attenuated inversion recovery; MD, mean diffusivity; UCSF-PDGM, University of California, San Francisco Preoperative Diffuse Glioma MRI; UPENN-GBM, glioblastoma patients from the University of Pennsylvania Health System.

Supplementary Figure 3.

(a) Scatter plot of MD and FW-corrected MD and (b) Bland–Altman plot showing the comparison of the MD and FW-corrected MD obtained from TrioTim in the UPENN-GBM dataset. The two lines indicate the 95% limit of agreement of CE ROI data.

CE, contrast enhanced; FLAIR, fluid-attenuated inversion recovery; FW, free water; MD, mean diffusivity; UPENN-GBM, glioblastoma patients from the University of Pennsylvania Health System.