Time-dependent Diffusion MRI–based Microstructural Mapping Value for Predicting Meningioma Grade, Subtype, and Proliferative Activity

Abstract

Purpose

To explore the diagnostic performance of time-dependent diffusion MRI (t_d-dMRI) combined with histogram analysis in predicting meningioma grade, subtype, and proliferative activity.

Materials and Methods

A total of 107 participants were prospectively enrolled (mean age ± SD, 55.7 years ± 10.2; 78 female) with meningiomas who underwent presurgical t_d-dMRI between August 2023 and July 2025. Histogram features from t_d-dMRI–derived parameters and semantic features from conventional MRI (cMRI) were estimated. Area under the receiver operating characteristic curve (AUC) and DeLong and integrated discrimination improvement (IDI) tests were used to evaluate model performance. Spearman rank correlations between t_d-dMRI metrics and the Ki-67 index were evaluated.

Results

In meningioma grading, the cMRI–t_d-dMRI combined model was superior to cMRI and single-parameter models (AUC, 0.86 [95% CI: 0.78, 0.92] vs 0.76 [95% CI: 0.67, 0.84] vs 0.64–0.71; corrected P = .009, .002–.009) and similar to the t_d-dMRI model (AUC, 0.83 [95% CI: 0.74, 0.89]; corrected P = .47). The combined model achieved a higher IDI than the cMRI and single-parameter models (0.16–0.27, all corrected P < .001). In meningioma subtyping, the combined model achieved the highest diagnostic performance (AUC, 0.86 [95% CI: 0.77, 0.93]) and a higher IDI than cMRI and single-parameter models (AUC, 0.14–0.31, all corrected P < .01). Weak correlations were observed between cell diameter, intracellular volume fraction, and cellularity and the Ki-67 index (r = −0.231 to −0.194 and 0.203–0.342; all P < .05).

Conclusion

The use of t_d-dMRI combined with histogram analysis performed well in assessing meningioma grade, subtype, and proliferative activity.

Keywords: CNS, MRI, MR-Diffusion Weighted Imaging, Time-dependent Diffusion MRI, Meningioma, Histogram Analysis, World Health Organization Grading, Histological Subtyping

Supplemental material is available for this article.

© RSNA, 2026

Summary

Histogram analysis of time-dependent diffusion MRI performed well in noninvasively assessing meningioma grade, subtype, and proliferative activity and could therefore be used to guide the development of personalized treatment strategies.

Key Points

• ■ In this prospective study of 107 participants with meningioma, a conventional MRI (cMRI)–time-dependent diffusion MRI (t_d-dMRI) combined model achieved the highest diagnostic performance (area under the receiver operating characteristic curve [AUC], 0.86) in differentiating low-grade and high-grade meningiomas (cMRI model: AUC, 0.76; single-parameter models: AUC, 0.64–0.71; t_d-dMRI model: AUC, 0.83; corrected P = .002–.47).

• ■ This cMRI–t_d-dMRI combined model also achieved the highest diagnostic performance (AUC, 0.86) in distinguishing fibrous from nonfibrous meningiomas (cMRI model: AUC, 0.73; single-parameter models: AUC, 0.71–0.85; t_d-dMRI model: AUC, 0.85; corrected P = .003–.81).

• ■ The Ki-67 index showed a weak correlation with cell diameter, intracellular volume fraction, and cellularity (r = −0.231 to −0.194 and 0.203–0.342; all P < .05).

Introduction

Meningiomas are the most common primary intracranial tumors and are categorized into three grades and 15 subtypes by the 2021 World Health Organization classification (1,2). The heterogeneity of meningioma grades and subtypes critically influences their growth patterns and recurrence rates. Most meningiomas are low-grade meningiomas (LGMs, grade 1), exhibiting indolent biologic behavior and a favorable prognosis (3). Nevertheless, approximately 20% of meningiomas are classified as high-grade meningiomas (HGMs, grades 2 and 3), demonstrating more aggressive behavior and higher recurrence rates (4,5). In asymptomatic patients with LGMs, an observation strategy may be a viable option (6). In contrast, if preoperative evaluation raises suspicion of HGMs, aggressive intervention (surgery and radiation therapy) should be pursued regardless of tumor size or symptom status. Notably, among LGMs, fibrous meningiomas usually show a characteristically firm consistency that complicates surgical resection compared with other LGMs. In addition, fibrous meningiomas are characterized by a high prevalence of chromosome 22q abnormalities and NF2 mutations (7,8). These factors may not only result in increased recurrence rates and intraoperative hemorrhage but also necessitate careful surgical strategy selection (the choice of surgical instrumentation, intraoperative transfusion planning, and allocation of longer operative time) (7,8). Therefore, preoperative grading and subtyping of meningiomas are clinically essential for therapeutic decision-making and prognostic evaluation.

Conventional MRI (cMRI) based on semantic features is the primary diagnostic tool for meningiomas. However, these cMRI features are highly observer-dependent and show overlap among different tumor grades (necrosis, peritumoral edema, or unclear tumor–brain interface) and subtypes (tumor location), resulting in insufficient diagnostic performance (4,9). Diffusion MRI techniques, including diffusion-weighted imaging based on pulsed gradient spin-echo (10), diffusion tensor imaging (11), and diffusion kurtosis imaging (12), have been extensively investigated for meningioma grading and subtyping. Although diffusion MRI can noninvasively assess the degree of mobility of water molecules within biologic tissue, conflicting findings on its utility in histologic classification have been reported (12–14). This limitation may arise because diffusion MRI cannot directly provide pathology features linked to tumor aggressiveness and subtype heterogeneity, such as cell size, cellularity, and nuclear size.

Recent advances in time-dependent diffusion MRI (t_d-dMRI) have revealed its unique advantages in characterizing tumor pathology in vivo (15,16). t_d-dMRI combines the oscillating gradient spin-echo method and pulsed gradient spin-echo method, which can explore the diffusion time dependence of restricted water diffusion (17). Additionally, the imaging microstructural parameters using limited spectrally edited diffusion method is a specific modeling approach for t_d-dMRI data. This method has been developed to quantify microstructural properties, such as cell diameter (d), intracellular volume fraction (f_in), and cellularity (16,18). Accumulating evidence has shown that t_d-dMRI–derived metrics correlate with key pathology features, including tumor grade, molecular subtypes, and Ki-67 proliferation indexes across various cancers, such as gliomas (19) and head and neck (20), breast (16), and prostate (15) cancers. Nevertheless, the clinical value of t_d-dMRI in imaging meningiomas remains to be elucidated.

We hypothesized that t_d-dMRI could better assess the histopathologic microstructure of meningiomas. The aims of this study were to investigate diagnostic performance of t_d-dMRI combined with histogram analysis in grading and subtyping meningiomas and to explore the correlation between t_d-dMRI metrics and the Ki-67 index.

Materials and Methods

One author (X.Z.), an MRI research scientist affiliated with Philips Healthineers, offered technical assistance (ie, brain t_d-dMRI sequence debugging) for this study under Philips collaborative regulations. This assistance was provided without remuneration or personal interests tied to the study. Another author (D.S.), who is not an employee of or consultant for Philips Healthineers, had control of the inclusion of any data and information that might present a conflict of interest for the other author (X.Z.). This prospective and preliminary study was approved by the local institutional review board ([2021] no. 580). Informed consent was obtained from all participants.

Study Participants

For this single-center prospective study, we recruited 270 consecutive patients with suspected intracranial extracerebral tumors at a tertiary care university hospital between August 2023 and July 2025. The inclusion criteria were as follows: participants underwent presurgical cranial MRI, including t_d-dMRI and conventional scanning, meningiomas were confirmed through histopathologic analysis based on the 2021 World Health Organization criteria, and participants underwent surgical intervention within 14 days of MRI. The exclusion criteria were as follows: insufficient image quality due to motion or susceptibility artifacts (n = 3), lesions less than 10 mm in maximum diameter (n = 1), inadequate histopathologic data (n = 5), and history of tumor-related treatments, including surgery, chemotherapy, or radiation therapy (n = 5). Finally, 107 patients (78 female and 29 male; mean age ± SD, 55.7 years ± 10.2 [range, 28–81 years]) were enrolled in this study. The participant enrollment study flowchart is shown in Figure 1.

Figure 1:

Flowchart shows participant enrollment. t_d-dMRI = time-dependent diffusion MRI.

Data Acquisition at MRI

All MRI examinations were performed on a 3.0-T scanner (Ingenia; Philips Healthcare) with a maximum gradient of 45 mT/m, a slew rate of 200 mT/m/sec, and a 32-channel head coil. t_d-dMRI was performed using an in-house–developed oscillating gradient spin-echo sequence with a trapezoid cosine gradient and pulsed gradient spin-echo sequence (21). Oscillating gradient spin-echo data were acquired at 17 Hz (effective diffusion time, 14.7 msec; one cycle; b = 0, 250, 500, 750, and 1000 sec/mm²) and 33 Hz (effective diffusion time, 7.6 msec; two cycles; b = 0, 100, 200, and 300 sec/mm²). Pulsed gradient spin-echo data were acquired with diffusion duration/separation of 86.3/12.0 msec at b values of 0, 300, 600, 900, 1200, 1500, and 1800 sec/mm². The following parameters were used for both sequences: repetition time/echo time, 3000/110 msec; field of view, 220 × 220 mm; matrix size, 112 × 109; sampling resolution, 2 × 2 mm²; section number, 22; section thickness, 5 mm; fat saturation method of spectral presaturation with inversion recovery; and sense factor, 3. The t_d-dMRI protocol scanning time was 8 minutes 24 seconds. The routine sequences included axial T2-weighted images, T2-fluid attenuated inversion recovery images, and unenhanced and enhanced T1-weighted images. For enhanced T1-weighted images, gadobenate dimeglumine (MultiHance; Bracco Diagnostics) was injected intravenously at a dose of 0.1 mmol/kg body weight and a flow rate of 2 mL/sec. Detailed imaging parameters are provided in Table S1.

Image Analysis

The t_d-diffusion data were fitted to a two-compartment model comprising impermeable spheres, in accordance with the imaging microstructural parameters, using the limited spectrally edited diffusion method (18,22). The fitted microstructural parameters (d, f_in, extracellular diffusivity [D_ex], and cellularity) were estimated using the following formulas:

$$\begin{array}{c}S=f_{in}\times S_{in}+\left(1-f_{in}\right)\times S_{ex},\\ S_{ex}=e^{-b\times D_{ex}},\\ cellularity=\frac{f_{in}}{d}\times 100,\end{array}$$

where S_ex is the D_ex MRI signal and S_in is the intracellular diffusion MRI signal (15,22). During the fitting process, the intracellular diffusivity was fixed at 2.0 µm²/msec because the fitting results were insensitive to variations in this parameter (23). The t_d-diffusion data were also fitted using the monoexponential model for apparent diffusion coefficient (ADC) calculation at each diffusion time (ADC measurement at 0 Hz [D_0Hz], ADC measurement at 17 Hz [D_17Hz], and ADC measurement at 33 Hz [D_33Hz]) using the following equation (24):

$$\frac{S}{S_0}=e^{-b\times ADC,}$$

where s₀ is the signal intensity in diffusion-weighted imaging with a b value of 0 sec/mm². Additionally, the relative ADC change at 17 Hz (ΔADC_17Hz) and the relative ADC change at 33 Hz (ΔADC_33Hz) were calculated using the following equation (20):

$${\text{ΔADC}}_{17Hz}=100\times \frac{D_{17Hz}-D_{0Hz}}{D_{0Hz}+eps}$$

where eps is a negligible constant added to the denominator to avoid division by zero. The fitting was performed using least-squares curve fitting in MATLAB 2022a (MathWorks).

Two experienced radiologists (H.Z. and D.J., with 9 and 4 years of experience in neuroradiology, respectively), blinded to participant data, independently performed manual tumor segmentation (ITK-SNAP software, version 3.8.0; http://www.itksnap.org). With cMRI as the reference, the regions of interest were manually delineated section by section along solid tumor margins on the b = 0 sec/mm² images. Necrotic and cystic areas, identified as hyperintensity on T2-weighted images and nonenhancing regions on enhanced T1-weighted images, were carefully excluded from the regions of interest (Fig S1). Histogram analyses of the above nine parameters were conducted, and 12 metrics were computed to provide quantitative assessment within regions of interest, including minimum, maximum, mean, and fifth (P5), 10th (P10), 25th (P25), 50th (P50), 75th (P75), 90th (P90), and 95th (P95) percentiles, skewness, and kurtosis. Intraclass correlation coefficients were calculated to confirm measurement reproducibility. Intraclass correlation coefficients greater than 0.75 indicated good agreement.

cMRI Semantic Feature Assessment

The cMRI semantic features were evaluated by a radiologist (D.S., with 15 years of experience in neuroradiology), who was blinded to the histopathologic results. These features included tumor size (longest diameter of tumor), location (non–skull base or skull base), shape (regular or irregular), base (broad base or narrow base), cyst, necrosis, or hemorrhage (no or yes), enhancement pattern (homogeneous or inhomogeneous), tumor–brain interface (clear or unclear), dural tail sign (no or yes), peritumoral edema (no or yes), midline shift (no or yes), and bone invasion (no or yes). Detailed definitions are provided in Table S2.

Histopathology Analysis

Histopathologic evaluation of hematoxylin-eosin staining was performed by a neuropathologist (Yu Zhang, MD, with 10 years of experience), who was blinded to the MRI findings. Immunohistochemical staining for Ki-67 was performed using a monoclonal antibody (MIB-1; Santa Cruz Biotechnology) (25). Ki-67 proliferative index quantification was performed by assessing the ratio of antibody-stained positive nuclei to total tumor nuclei. The highest value from five representative high-power fields was recorded.

Statistical Analysis

The null hypothesis of an area under the receiver operating characteristic curve (AUC) was 0.50, and the alternative hypothesis was an AUC of 0.75; the statistical power was 90%; and the average proportions were 80% and 20% for LGMs and HGMs, respectively. The minimum required sample sizes were thus 68 LGMs and 17 HGMs. Continuous variables were reported as medians with IQRs and categorical variables were reported as proportions with percentages. We performed group comparisons using the Student t test or Mann-Whitney U test for continuous variables and the χ² test for categorical variables.

For each predictive model, we first implemented the least absolute shrinkage and selection operator with 10-fold cross-validation to select variables from all variables. The 10-fold cross-validation was used to find an optimal regularization parameter (λ) in the least absolute shrinkage and a selection operator that minimized the binomial deviance. We applied the optimal λ + 1 standard error to select the variables with nonzero coefficients. Second, we assessed the multicollinearity among the selected features using the variance inflation factor analysis. When the variance inflation factor was above 10 (26), the feature was eliminated. Then, we applied logistic regression to build predictive models using the selected features, including the cMRI model, single-parameter model, t_d-dMRI model, and combined model.

We used the Hosmer-Lemeshow test to evaluate the goodness of fit of the logistic regression models. We used AUC, specificity, sensitivity, accuracy, and 95% CIs to evaluate the diagnostic performance by receiver operating characteristic curve analysis. We used the DeLong test and integrated discrimination improvement (IDI) to compare model performance, incorporating false discovery rate correction for multiple comparisons. We performed bootstrap resampling (n = 1000) to assess the stability of the combined model’s performance for each sample.

We performed Spearman rank correlation to assess the correlations between t_d-dMRI metrics and the Ki-67 index. We considered a two-tailed P value less than .05 to indicate statistical significance. One author (H.Z., with 9 years of experience) performed statistical analysis using SPSS software (version 21.0; IBM), R software (version 4.2.2; R Project for Statistical Computing), MedCalc (version 15.2.2; MedCalc Software), and GraphPad Prism (version 8.0.1; GraphPad Software).

Results

Among the 107 enrolled participants, 85 of 107 (79.4%) participants were diagnosed with LGMs (grade 1): 26 fibrous meningiomas, 23 meningothelial meningiomas, 25 transitional meningiomas, four secretory meningiomas, four angiomatous meningiomas, one psammomatous meningioma, and two unclassified meningiomas. The remaining 22 of 107 (20.6%) participants were confirmed to have HGMs (grade 2–3): 20 atypical meningiomas, one chordoid meningioma, and one anaplastic meningioma.

Comparisons of Demographic and cMRI Characteristics between HGM and LGM Groups

The differential distributions of demographic and cMRI characteristics between HGMs and LGMs are shown in Table 1. Compared with LGMs, HGMs demonstrated larger tumor size (median, 4.82 cm [IQR, 3.92–6.52 cm] vs median, 3.62 cm [IQR, 2.21–4.55 cm]; P < .001) and a higher proportion of irregular shape (17 of 22 [77%] vs 34 of 85 [40%]; P = .002), cyst, necrosis, or hemorrhage (10 of 22 [45%] vs 12 of 85 [14%]; P = .003), inhomogeneous enhancement (15 of 22 [68%] vs 33 of 85 [39%]; P = .01), unclear tumor–brain interface (six of 22 [27%] vs five of 85 [6%]; P = .01), peritumoral edema (19 of 22 [86%] vs 37 of 85 [44%]; P < .001), and bone invasion (10 of 22 [45%] vs 20 of 85 [24%]; P = .04). We found no evidence of intergroup differences in age (mean ± SD, 55.1 years ± 10.6 vs 58.0 years ± 8.4; P = .24), sex (female, 64 of 85 [75%]) vs 14 of 22 [64%]; P = .27), location (skull base, 39 of 85 [75%]) vs 14 of 22 [64%]; P = .27), broad base (81 of 85 [95%]) vs 20 of 22 [91%]; P = .78), or dural tail sign (81 of 85 [95%]) vs 21 of 22 [95%]; P > .99).

Table 1:

Comparisons of Participant Demographic and Conventional MRI Semantic Characteristics between the LGM and HGM Groups

Characteristic	Total (n = 107)	LGMs (n = 85)	HGMs (n = 22)	P Value
Age (y)	55.7 ± 10.2 (28.00–81.00)	55.1 ± 10.6 (28.00–81.00)	58.0 ± 8.4 (44.00–75.00)	.24*
Sex
Male	29 (27.1)	21 (25)	8 (36)	.27†
Female	78 (72.9)	64 (75)	14 (64)
Ki-67 index (%)‡	5.00 (2.00–5.00)	3.00 (2.00–5.00)	5.00 (5.00–8.00)	<.001‡
cMRI semantic features
Tumor size (cm)	3.95 (2.39–4.95)	3.62 (2.21–4.55)	4.82 (3.92–6.52)	<.001‡
Location				.68†
Non–skull base	59 (55.1)	46 (54)	13 (59)
Skull base	48 (44.9)	39 (46)	9 (41)
Shape				.002†
Regular	56 (52.3)	51 (60)	5 (23)
Irregular	51 (47.7)	34 (40)	17 (77)
Base				.78†
Broad base	101 (94.4)	81 (95)	20 (91)
Narrow base	6 (5.6)	4 (5)	2 (9)
Cyst, necrosis, or hemorrhage				.003†
No	85 (79.4)	73 (86)	12 (55)
Yes	22 (20.6)	12 (14)	10 (45)
Enhancement pattern				.01†
Homogeneous	59 (55.1)	52 (61)	7 (32)
Inhomogeneous	48 (44.9)	33 (39)	15 (68)
Tumor–brain interface				.01†
Clear	96 (89.7)	80 (94)	16 (73)
Unclear	11 (10.3)	5 (6)	6 (27)
Dural tail sign				>.99†
No	5 (4.8)	4 (5)	1 (5)
Yes	102 (95.3)	81 (95)	21 (95)
Peritumoral edema				<.001†
No	51 (47.7)	48 (56)	3 (14)
Yes	56 (52.3)	37 (44)	19 (86)
Midline shift				.06†
No	76 (71.0)	64 (75)	12 (55)
Yes	31 (29.0)	21 (25)	10 (45)
Bone invasion				.04†
No	77 (72.0)	65 (76)	12 (55)
Yes	30 (28.0)	20 (24)	10 (45)

Note.—Data are means ± SDs with IQRs in parentheses, medians with IQRs in parentheses, or numbers with percentages in parentheses. cMRI = conventional MRI, HGM = high-grade meningioma, LGM = low-grade meningioma.

^*Data comparisons were performed with the Student t test.

^† Data comparisons were performed with the χ² test.

^‡ Data comparisons were performed with the Mann-Whitney U test.

Interobserver Concordance

Good interobserver agreement was obtained for 105 t_d-dMRI histogram metrics, with intraclass correlation coefficients ranging from 0.77 to 0.99, excluding D_ex_kurtosis, D_0Hz_kurtosis, and D_33Hz_kurtosis (Table S3). Therefore, in the subsequent analysis, we used the remaining 105 metrics obtained by the senior radiologist.

Comparisons of t_d-dMRI Metrics between HGMs and LGMs

The histogram features of t_d-dMRI metrics with significant differences for meningioma grading are listed in Table 2. The complete data are presented in Table S4. Compared with LGMs, HGMs showed higher values for the following parameters (all P < .05): f_in (mean, P50, P75, P90, and P95), D_ex (max, mean, P50, P75, and P90), cellularity (max, P75, P90, and P95), D_0Hz (max), D_17Hz (P95), ΔADC_17Hz (max, mean, P50, P75, P90, and P95), and ΔADC_33Hz (max, mean, P50, and P75). Notably, among these, the higher-percentile metrics (P75, P90, P95, and max) from D_ex, D_0Hz, and D_17Hz were consistently elevated in HGMs. The distribution patterns of six representative parameters in LGM and HGM groups are shown in Figure 2, and metrics and images from representative participants in the LGM and HGM groups are provided in Figure 3.

Table 2:

Comparisons of Time-dependent Diffusion MRI Parameters with Significant Differences between LGM and HGM Groups

Variable	LGMs (n = 85)	HGMs (n = 22)	P Value
fin
fin_mean	0.22 (0.19–0.26)	0.25 (0.21–0.28)	.03
fin_P50	0.22 (0.20–0.26)	0.26 (0.22–0.30)	.02
fin_P75	0.26 (0.23–0.29)	0.29 (0.27–0.33)	.009
fin_P90	0.30 (0.26–0.33)	0.33 (0.29–0.36)	.01
fin_P95	0.32 (0.27–0.36)	0.35 (0.32–0.38)	.02
Dex (μm2/msec)
Dex_max	3.00 (2.60–3.00)	3.00 (3.00–3.00)	.002
Dex_mean	1.40 (1.25–1.72)	1.55 (1.44–1.93)	.02
Dex_P50	1.39 (1.25–1.70)	1.56 (1.38–1.89)	.03
Dex_P75	1.52 (1.39–2.02)	1.86 (1.58–2.18)	.03
Dex_P90	1.81 (1.54–2.28)	2.13 (1.84–2.69)	.02
Cellularity (μm−1)
Cellularity_max	5.81 (5.23–6.71)	6.65 (5.90–7.43)	.005
Cellularity_P75	3.64 (3.27–4.03)	4.15 (3.70–4.75)	.01
Cellularity_P90	4.29 (3.85–4.84)	4.89 (4.25–5.41)	.02
Cellularity_P95	4.61 (4.16–5.29)	5.26 (4.58–5.76)	.03
D0Hz (μm2/msec)
D0Hz_max	1.30 (1.05–1.73)	1.59 (1.33–1.82)	.02
D17Hz (μm2/msec)
D17Hz_P95	1.10 (1.01–1.33)	1.25 (1.13–1.57)	.04
ΔADC17Hz (%)
ΔADC17Hz_max	34.42 (28.35–41.46)	46.26 (29.39–59.37)	.02
ΔADC17Hz_mean	11.89 (10.24–13.91)	14.14 (12.15–15.00)	.008
ΔADC17Hz_P50	11.31 (9.81–13.87)	13.45 (11.40–14.94)	.02
ΔADC17Hz_P75	15.49 (13.13–17.10)	17.09 (15.55–18.84)	.01
ΔADC17Hz_P90	19.33 (16.31–21.53)	21.65 (18.50–25.33)	.02
ΔADC17Hz_P95	21.60 (18.74–24.61)	23.88 (21.25–28.93)	.03
ΔADC33Hz (%)
ΔADC33Hz_max	65.24 (56.27–79.03)	73.43 (65.48–150.45)	.03
ΔADC33Hz_mean	32.98 (28.00–40.52)	38.58 (32.45–43.10)	.045
ΔADC33Hz_P50	34.15 (28.04–40.04)	39.04 (32.25–43.84)	.03
ΔADC33Hz_P75	39.64 (34.32–47.00)	44.11 (38.43–49.42)	.04

Note.—Data are presented as medians with IQRs in parentheses. Data comparisons were performed with the Mann-Whitney U test. ADC = apparent diffusion coefficient, D_ex = extracellular diffusivity, D_0Hz = ADC measurement at 0 Hz, D_17Hz = ADC measurement at 17 Hz, f_in = intracellular volume fraction, HGM = high-grade meningioma, LGM = low-grade meningioma, max = maximum, P50 = 50th percentile, P75 = 75th percentile, P90 = 90th percentile, P95 = 95th percentile, ΔADC_17Hz = relative ADC change at 17 Hz, ΔADC_33Hz = relative ADC change at 33 Hz.

Figure 2:

The violin plots show the distribution of useful time-dependent diffusion MRI metrics for differentiating (A) low-grade meningiomas (LGMs) from high-grade meningiomas (HGMs) and (B) nonfibrous meningiomas (non-Fibs) from fibrous meningiomas (Fibs). The three dashed lines in each violin plot mark the median (center) and the 25th and 75th percentiles (the IQR). *P < .05, **P < .01, ***P < .001. Data comparisons were performed with the Mann-Whitney U test. ADC = apparent diffusion coefficient, f_in = intracellular volume fraction, D_ex = extracellular diffusivity, D_0Hz = ADC measurement at 0 Hz, D_17Hz = ADC measurement at 17 Hz, D_33Hz = ADC measurement at 33 Hz, max = maximum, P75 = 75th percentile, P90 = 90th percentile, P95 = 95th percentile, ΔADC_17Hz = relative ADC change at 17 Hz, ΔADC_33Hz = relative ADC change at 33 Hz.

Figure 3:

Images show representative time-dependent diffusion MRI (t_d-dMRI) metrics and corresponding axial T2-weighted images (T2WI) and hematoxylin-eosin (H&E) staining of meningiomas from participants who underwent contrast-enhanced MRI. (A) Images in a 61-year-old female patient with fibrous meningioma, (B) images in a 57-year-old male patient with transitional meningioma, and (C) images in a 52-year-old female patient with atypical meningioma. (A, B) The conventional MRI (cMRI) model misclassified both low-grade meningiomas (LGMs) as high-grade meningiomas (HGMs) (predictive scores, −1.07 and 0.01 for meningiomas in panels A and B, respectively), whereas the combined model correctly identified the meningioma in panel A as an LGM (score, −2.44) but still misclassified the meningioma in panel B as an HGM (score, 0.31). (C) Both the cMRI and combined models confirmed this meningioma as an HGM (scores, 0.01 and 0.67, respectively). (A) The cMRI model misclassified this fibrous meningioma as a nonfibrous meningioma (score, −1.75), whereas the combined model correctly classified it as fibrous meningioma (score, −0.41). (B) The cMRI model misclassified this nonfibrous meningioma as a fibrous meningioma (score, −0.36), whereas the combined model correctly classified it as a nonfibrous meningioma (score, −1.61). ADC = apparent diffusion coefficient, d = cell diameter, D_ex = extracellular diffusivity, D_0Hz = ADC measurement at 0 Hz, D_17Hz = ADC measurement at 17 Hz, D_33Hz = ADC measurement at 33 Hz, f_in = intracellular volume fraction, IMPULSED = imaging microstructural parameters using limited spectrally edited diffusion, ΔADC_17Hz = relative ADC change at 17 Hz, ΔADC_33Hz = relative ADC change at 33 Hz.

Diagnostic Performance of Predictive Models in Meningioma Grading

Receiver operating characteristic curve analysis metrics for the logistic models are presented in Tables 3 and S2 and Figure 4. In the cMRI model, cyst, necrosis, or hemorrhage and peritumoral edema were identified as strong predictors of HGMs (AUC, 0.76 [95% CI: 0.67, 0.84]). In the t_d-dMRI model, 10 parameters (f_in_max, D_ex_max, D_ex_P50, cellularity_min, cellularity_P75, cellularity_kurtosis, D_0Hz_skewness, D_17Hz_P95, ΔADC_17Hz_P5, and ΔADC_33Hz_max) showed the best performance for differentiating HGMs from LGMs, achieving the highest AUC of 0.83 (95% CI: 0.74, 0.89) compared with the cMRI and single-parameter models (AUC, 0.64–0.76).

Table 3:

Logistic Regression Model Performance in the Prediction of Meningioma Grades

Model	Cutoff	AUC (95% CI)	Sensitivity (n = 22)	Specificity (n = 85)	Accuracy (n = 107)
cMRI model	−2.87	0.76 (0.67, 0.84)	21 (95)	47 (55)	68 (63.6)
fin model	−1.08	0.67 (0.57, 0.75)	11 (50)	68 (80)	79 (73.8)
Dex model	−1.44	0.71 (0.61, 0.79)	17 (77)	53 (62)	70 (65.4)
Cellularity model	−0.97	0.71 (0.62, 0.79)	11 (50)	73 (86)	84 (79.4)
D0Hz model	−1.13	0.65 (0.55, 0.74)	10 (45)	70 (82)	80 (74.8)
D17Hz model	−1.54	0.64 (0.54, 0.73)	17 (77)	47 (55)	64 (59.8)
ΔADC17Hz model	−1.59	0.67(0.57. 0.76)	20 (91)	33 (39)	53 (49.5)
ΔADC33Hz model	−1.51	0.66 (0.57, 0.75)	18 (82)	46 (54)	64 (59.8)
td-dMRI model	−0.80	0.83 (0.74, 0.89)	15 (68)	73 (86)	88 (82.2)
Combined model*	−1.94	0.86 (0.78, 0.92)	20 (91)	54 (64)	74 (69.2)

Note.—Data are numbers with percentages in parentheses unless otherwise specified. ADC = apparent diffusion coefficient, AUC = area under the receiver operating characteristic curve, cMRI = conventional MRI, D_ex = extracellular diffusivity, D_0Hz = ADC measurement at 0 Hz, D_17Hz = ADC measurement at 17 Hz, f_in = intracellular volume fraction, ΔADC_17Hz = relative ADC change at 17 Hz, ΔADC_33Hz = relative ADC change at 33 Hz, t_d-dMRI = time-dependent diffusion MRI.

^*The combined model consisted of the combination of the conventional MRI, f_in, D_ex, cellularity, D_0Hz, D_17Hz, ΔADC_17Hz, and ΔADC_33Hz models.

Figure 4:

Receiver operating characteristic curves show the performance of different logistic models for differentiating (A) low-grade meningiomas (LGMs) from high-grade meningiomas (HGMs) and (B) nonfibrous meningiomas from fibrous meningiomas. Data shown for logistic models are AUCs with 95% CIs in brackets. ADC = apparent diffusion coefficient, AUC = area under the receiver operating characteristic curve, cMRI = conventional MRI, D_ex = extracellular diffusivity, D_0Hz = ADC measurement at 0 Hz, D_17Hz = ADC measurement at 17 Hz, D_33Hz = ADC measurement at 33 Hz, f_in = intracellular volume fraction, ΔADC_17Hz = relative ADC change at 17 Hz, ΔADC_33Hz = relative ADC change at 33 Hz, t_d-dMRI = time-dependent diffusion MRI.

In the combined model, logistic regression analyses showed that cyst, necrosis, or hemorrhage and peritumoral edema from cMRI and f_in_P50, D_ex_max, D_ex_P50, cellularity_kurtosis, and D_17Hz_P95 from t_d-dMRI were strong predictors of meningioma grade. Hosmer-Lemeshow goodness-of-fit test results suggested that the combined model was well calibrated (P = .79). The combined model achieved the highest diagnostic performance (AUC, 0.86 [95% CI: 0.78, 0.92]; sensitivity, 91% [20 of 22]; specificity, 64% [54 of 85]). To assess combined model performance stability, bootstrapping with 1000 resamples showed a mean AUC of 0.86 (95% CI: 0.77, 0.93), a sensitivity of 82% (18 of 22; 95% CI: 56, 100), a specificity of 79% (67 of 85; 95% CI: 53, 97), and an accuracy of 79.4% (85 of 107; 95% CI: 61.7, 92.5).

As shown in Table 4, the DeLong tests revealed that the performance of the combined model was superior to that of the cMRI model and each of the single-parameter models (all corrected P < .05), whereas we found no evidence of a difference in performance between the combined model and the t_d-dMRI model (corrected P = .47). Of note, the combined model showed improved discrimination over the cMRI and single-parameter models with IDIs ranging from 0.16 to 0.27. As shown in Figure 5A, the cMRI model misclassified 39 meningiomas (38 LGMs and one HGM), whereas the combined model correctly graded 10 of these meningiomas (26% [10 of 39], all LGMs). The combined model also misclassified 33 meningiomas (31 LGMs and two HGMs), four of which (12% [four of 33], three LGMs, and one HGM) were correctly classified by the cMRI model.

Table 4:

Comparisons of Diagnostic Performance among Different Models for Predicting Meningioma Grades

Model*	DeLong Test Corrected P Value†	IDI (95% CI)	IDI Test Corrected P Value†
Combined model vs cMRI model	.009	0.16 (0.07, 0.24)	<.001
Combined model vs fin model	.003	0.25 (0.14, 0.36)	<.001
Combined model vs Dex model	.004	0.23 (0.12, 0.33)	<.001
Combined model vs Cellularity model	.003	0.21 (0.11, 0.31)	<.001
Combined model vs D0Hz model	.002	0.24 (0.14, 0.33)	<.001
Combined model vs D17Hz model	.002	0.26 (0.16, 0.37)	<.001
Combined model vs ΔADC17Hz model	.002	0.27 (0.17, 0.37)	<.001
Combined model vs ΔADC33Hz model	.002	0.27 (0.17, 0.36)	<.001
Combined model vs td-dMRI model	.47	0.04 (−0.04, 0.13)	.34

Note.—ADC = apparent diffusion coefficient, cMRI = conventional MRI, D_ex = extracellular diffusivity, D_0Hz = ADC measurement at 0 Hz, D_17Hz = ADC measurement at 17 Hz, f_in = intracellular volume fraction, IDI = integrated discrimination improvement, ΔADC_17Hz = relative ADC change at 17 Hz, ΔADC_33Hz = relative ADC change at 33 Hz, t_d-dMRI = time-dependent diffusion MRI.

^*Combined model consisted of the combination of the conventional MRI, f_in, D_ex, cellularity, D_0Hz, D_17Hz, ΔADC_17Hz, and ΔADC_33Hz models. False discovery rate correction was applied in the multiple comparisons.

^† Corrected P value is less than .05.

Figure 5:

Confusion matrices for conventional MRI (cMRI) and combined models show meningioma grading (A) and subtyping (B). (A) Among the 39 meningiomas misdiagnosed by the cMRI model (38 low-grade meningiomas [LGMs], one high-grade meningioma [HGM]), the combined model correctly classified 10 meningiomas (26% [10 of 39]; all LGMs). The combined model also misdiagnosed 33 meningiomas (31 LGMs, two HGMs), four of which (12% [four of 33]; three LGMs, one HGM) were correctly diagnosed by the cMRI model. (B) Among the 25 meningiomas misdiagnosed by the cMRI model (19 nonfibrous meningiomas, six fibrous meningiomas), the combined model correctly subtyped 16 meningiomas (64% [16 of 25]; 11 nonfibrous meningiomas, five fibrous meningiomas). The combined model also misdiagnosed 16 meningiomas (13 nonfibrous meningiomas, three fibrous meningiomas), seven of which (44% [seven of 16]; five nonfibrous meningiomas, two fibrous meningiomas) were correctly subtyped by the cMRI model. FN = false-negative, FP = false-positive, TN = true-negative, TP = true-positive.

Comparisons of Demographic and cMRI Characteristics and t_d-dMRI Metrics between the Fibrous Meningioma and Nonfibrous Meningioma Groups

The LGM group was further classified into 26 of 85 (31%) participants with fibrous meningiomas and 59 of 85 (69%) participants with nonfibrous meningiomas. The fibrous meningioma group had a higher proportion of participants with non–skull base involvement than the nonfibrous meningioma group (21 of 26 [80.77%] vs 25 of 59 [42.37%]; P = .001) (Table S6). We found no evidence of differences in age, sex, or other cMRI semantic features between the fibrous meningioma and nonfibrous meningioma groups (P = .06 to .98).

The histogram features of t_d-dMRI metrics for meningioma subtyping are listed in Table S7. Compared with the nonfibrous meningioma group, the fibrous meningioma group showed lower values of the following parameters (all P < .05): f_in (mean, P50, P75, P90, and P95), D_ex (P75, P90, and P95), cellularity (max, P75, P90, and P95), D_33Hz (max, P75, P90, and P95), ΔADC_17Hz (max, mean, P50, P75, P90, and P95), and ΔADC_33Hz (max, mean, P10, P25, P50, P75, P90, and P95). The fibrous meningioma group showed higher values of D_ex_min and D_ex_skewness than the nonfibrous meningioma group (all P < .05). The distribution patterns of six representative parameters in the fibrous meningioma and nonfibrous meningioma groups are shown in Figure 2, and metrics and images from representative participants in these two groups are provided in Figure 3.

Diagnostic Performance of Predictive Models in Meningioma Subtyping

Results of receiver operating characteristic curve analyses of the logistic regression models in meningioma subtyping are presented in Table S8 and Figure 4. The ΔADC_17Hz model achieved a higher AUC (0.85; 95% CI: 0.75, 0.91) than the cMRI model and other single-parameter models (f_in, D_ex, cellularity, D_33Hz, and ΔADC_33H; AUC, 0.71–0.79). In the combined model, logistic regression analyses showed that location and tumor size from cMRI and ΔADC_17Hz_P5, ΔADC_17Hz_P75, ΔADC_17Hz_P95, and ΔADC_17Hz_kurtosis from t_d-dMRI were strong predictors of meningioma subtype. The combined model achieved the highest diagnostic performance (AUC, 0.86 [95% CI: 0.77, 0.93]; sensitivity, 88% [23 of 26]; specificity, 78% [46 of 59]) with good calibration (P = .43). Bootstrapping with 1000 resamples showed a mean AUC of 0.86 (95% CI: 0.76, 0.94), a sensitivity of 85% (22 of 26; 95% CI: 71, 100), a specificity of 81% (48 of 59; 95% CI: 69, 93), and an accuracy of 84% (71 of 85; 95% CI: 74, 92) for the combined model.

The DeLong test revealed that the combined model performance was superior to that of the cMRI, D_ex, cellularity, and D_33Hz models (all corrected P < .05), whereas we found no difference in performance between the combined model and the rest of the single-parameter models or the t_d-dMRI model (Table S9). Of note, discrimination was improved with the combined model compared with the cMRI and single-parameter models (f_in, D_ex, cellularity, D_33Hz, and ΔADC_33H), with IDIs ranging from 0.14 to 0.31 (all corrected P < .01). As shown in Figure 5B, the cMRI model misdiagnosed 25 meningioma subtypes (19 nonfibrous meningiomas and six fibrous meningiomas), whereas the combined model correctly subtyped 16 of these meningiomas (64% [16 of 25]; 11 nonfibrous meningiomas and five fibrous meningiomas). The combined model also misdiagnosed 16 meningiomas (13 nonfibrous meningiomas and three fibrous meningiomas), seven of which (44% [seven of 16]; five nonfibrous meningiomas and two fibrous meningiomas) were correctly subtyped by the cMRI model.

Correlation of t_d-dMRI Metrics with the Ki-67 Index

Spearman correlations between all t_d-dMRI metrics and the Ki-67 index are presented in Table S10, and the correlations that reached statistical significance (P < .05) are highlighted in Figure 6. d_min, d_P25, d_P50, d_P75, and cellularity_skewness showed weak negative correlations with the Ki-67 index (r = −0.231 to −0.194; all P < .05). In contrast, f_in_mean, f_in_P25, f_in_P50, f_in_P75, f_in_P90, cellularity_max, cellularity_mean, cellularity_P25, cellularity_P50, cellularity_75, cellularity_90, and cellularity_95 showed weak positive correlations with the Ki-67 index (r = 0.203–0.342; all P < .05).

Figure 6:

The correlation matrix shows statistically significant correlations (all P < .05) between the Ki-67 index and time-dependent diffusion MRI (t_d-dMRI) metrics. d = cell diameter, f_in = intracellular volume fraction, min = minimum, max = maximum, P25 = 25th percentile, P50 = 50th percentile, P75 = 75th percentile, P90 = 90th percentile, P95 = 95th percentile.

Discussion

Although preoperative grade and subtype are linked to therapeutic decision-making and prognostic evaluation in meningiomas, its correlations with tumor microscopic properties, measuring with t_d-dMRI, remains unclear. Our results demonstrated that the cMRI–t_d-dMRI combined model demonstrated the best diagnostic performance in meningioma grade (AUC, 0.86 [95% CI: 0.78, 0.92]), with its AUC value surpassing that of cMRI and single-parameter models (P < .01 for all AUCs). The combined model also achieved the highest diagnostic performance in meningioma subtype (AUC, 0.86 [95% CI: 0.77, 0.93]). Moreover, the tumor microstructure-related parameters d, f_in, and cellularity were significantly correlated with the Ki-67 index (r = −0.231 to −0.194 and 0.203–0.342; all P < .05).

Preoperative prediction of HGMs with a high-confidence method is crucial for guiding aggressive surgical management. Although diffusion-weighted imaging based on pulsed gradient spin-echo was the first diffusion MRI used in clinical settings, the utility of the quantitative parameter ADC for differentiating meningioma grade remains controversial. Some studies reported that HGMs showed lower ADC values than LGMs (27,28), whereas others revealed higher ADC values in HGMs (10,29). t_d-dMRI, a novel diffusion technique capable of characterizing tissue microstructure, has been applied to extra-axial and intra-axial brain tumors (19,30,31) but has rarely been investigated for meningiomas. Our results demonstrated that HGMs exhibited higher D_ex, D_0Hz, and D_17Hz values (especially in high-percentile metrics), aligning with studies of elevated ADC (10,29). This discrepancy may be attributed to differences in inclusion subtypes, as well as the use of shorter diffusion time, longer b value range, and distinct region of interest strategies in our study. Whole-tumor histogram analysis used in our study may capture intratumoral micronecrosis and cystic degeneration in HGMs more accurately than local region of interest methods (3,32). In our study, several ΔADC_17Hz and ΔADC_33Hz histogram metrics were also higher in HGMs than LGMs. The relative ADC change values offer greater reliability than ADC values by capturing the stronger diffusion time dependence in HGMs with abundant microstructural barriers. Beyond diffusion metrics, microstructural parameters derived from t_d-dMRI (f_in and cellularity) were higher in HGMs and showed positive correlations with the Ki-67 index, whereas the d value showed a negative correlation. The results were consistent with pathologic features of HGMs, which are characterized by high mitotic activity, prominent nucleoli, and small cells with a high nuclear–cytoplasmic ratio (2).

The cMRI–t_d-dMRI combined model achieved the highest diagnostic performance (AUC, 0.86 [95% CI: 0.78, 0.92]). This observed effect size exceeded the anticipated effect size used in the power analysis (AUC, 0.75), suggesting that the diagnostic signal detected by t_d-dMRI is stronger than initially anticipated and clinically meaningful. The combined model outperformed the cMRI and single-parameter models (IDI, 0.16–0.27; all P < .001) and showed performance comparable to radiomics-based study or imaging score (AUC, 0.82–0.86) (9,33). Notably, the combined model correctly classified 26% (10 of 39) of meningiomas misclassified by the cMRI model, whereas the cMRI model correctly classified only 12% (four of 33) of meningiomas misclassified by the combined model. These findings highlight the potential of t_d-dMRI as an imaging method for meningioma grading and as a complementary tool that significantly enhances diagnostic performance when integrated with cMRI.

Preoperative identification of fibrous subtypes is crucial because these firmer tumors require different surgical techniques and longer operative times. Our study demonstrated that histogram metrics derived from t_d-dMRI can differentiate meningioma subtypes. Specifically, fibrous meningiomas exhibited significantly lower f_in, cellularity, D_ex, and D_33Hz values, reflecting diffusion restriction and sparse cellularity due to abundant collagen deposition and vitreous degeneration (34). We found no evidence of differences in D_0Hz values, consistent with previous ADC assessments (8,10), indicating that long diffusion time (pulsed gradient spin-echo) is less sensitive to smaller spatial scales than short diffusion time (oscillating gradient spin-echo) for subtyping. Additionally, nonfibrous meningiomas showed higher ΔADC_17Hz and ΔADC_33Hz values, reflecting stronger diffusion time dependence. This is attributed to the predominance of meningothelial and transitional subtypes (81% [69 of 85]), whose higher cellularity and lobular architecture (34,35), creating more diffusion barriers that amplify greater time-dependent signal change. Among predictive models, the combined model achieved the highest performance (AUC, 0.86 [95% CI: 0.77, 0.93]), with relative ADC change values as major predictors. The combined model demonstrated improvement over the cMRI and single-parameter (f_in, D_ex, cellularity, D_33Hz, and ΔADC_33Hz) models (IDI, 0.14–0.31; all corrected P < .01), and correctly classified 64.00% (16 of 25) of meningiomas misclassified by the cMRI model. These findings underscore that t_d-dMRI provides complementary information to cMRI, enabling the combined model to identify diffusion restriction at a smaller spatial scale.

Notably, the t_d-dMRI sequence in our study was acquired on a clinically available 3.0-T MRI scanner. Although the scanning time (8 minutes 24 seconds) is longer than that of diffusion-weighted imaging, it remains clinically feasible. Our findings demonstrate that the t_d-dMRI metrics are biophysically supported and explainable in meningioma characterization. The thresholds suggested by our results may prompt more aggressive resection and immediate postoperative radiation therapy. However, these thresholds are preliminary, and future validation in larger samples is warranted.

Our study has limitations. First, the limited sample sizes of HGMs and fibrous meningiomas relative to the number of predictors may cause overfitting despite our use of the least absolute shrinkage and selection operator and cross-validation in our study. External validation of our preliminary findings in larger, multicenter samples is required. Second, although we explored the correlation between t_d-dMRI histogram metrics and the Ki-67 index, future studies should include histopathology validation through precise imaging–pathology correlation. Third, necrotic and cystic regions were excluded during region of interest delineation. Because the model was trained only on the solid tumor components, it may not fully represent the heterogeneity of the entire lesion. Finally, semantic features were assessed by a single radiologist and therefore may be subjective owing to a lack of interobserver validation. Multireader evaluations are needed in future studies.

In conclusion, histogram analysis of t_d-dMRI is a noninvasive and promising approach to evaluate grades and subtypes of meningiomas. In addition, t_d-dMRI metrics modestly correlate with the Ki-67 index, warranting further investigation. Among predictive models, the cMRI–t_d-dMRI combined model showed the greatest potential in meningioma grading and subtyping. Future studies can focus on integrating t_d-dMRI with radiomics or deep learning to improve predictive performance and enhance tumor characterization.

Disclosures of conflicts of interest

Please see ICMJE form(s) for author conflicts of interest. These have been provided as supplemental materials.