Habitat subregions analysis based on neurite orientation dispersion and density imaging enhances isocitrate dehydrogenase genotyping in glioma

Abstract

Background

Noninvasive isocitrate dehydrogenase (IDH) genotyping in gliomas remains a critical challenge. This study investigates the performance of the whole-tumor histogram analysis of neurite orientation dispersion and density imaging (NODDI) and diffusion tensor imaging (DTI) in IDH genotyping and further explores their differences across habitat subregions.

Methods

This prospective study enrolled participants with suspected gliomas who underwent MRI scans before surgery and calculated diffusion metrics from DTI and NODDI. The whole-tumor region, including tumors and peritumoral edema, was delineated. Otsu’s thresholding method was used to divide the whole-tumor region into Habitat D (DTI-based, Otsu-segmented) based on fractional anisotropy (FA) and mean diffusivity (MD) derived from DTI, and into Habitat N (NODDI-based, Otsu-segmented) based on intracellular volume fraction (ICVF) and orientation dispersion index (ODI) derived from NODDI. Histogram features were extracted from the whole-tumor region and each habitat’s subregions. The Mann-Whitney U test was used to assess the differences in histogram features between different IDH genotypes. Logistic regression models were established to predict IDH genotypes. ROC curve analysis and DeLong tests were employed to evaluate and compare the diagnostic performance.

Results

A total of 75 participants with IDH-wildtype (n = 39) and IDH-mutant (n = 36) glioma were included. In the whole-tumor region, NODDI and DTI showed comparable diagnostic performance in IDH genotyping (AUC = 0.858 and 0.788, respectively; p > 0.05). In the habitat subregions, the histogram features in the Habitat N enhance IDH genotyping performance compared to the whole-tumor region, with the NODDI model outperforming the DTI model (AUC = 0.944 and 0.863, respectively; p < 0.05). The nomogram integrating age and the optimal NODDI model achieved high diagnostic performance (AUC = 0.962).

Conclusions

NODDI-based habitat subregions analysis is a promising approach to further enhance the diagnostic performance of DTI and NODDI histogram features in glioma IDH genotyping, and to capitalize on the advantages of NODDI in capturing the heterogeneity of microstructure.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-025-01934-4.

Introduction

Diffuse adult-type gliomas represent the most prevalent primary malignant brain neoplasms [1]. In the fifth edition of the World Health Organization (WHO) Classification of Tumors of the Central Nervous System, isocitrate dehydrogenase (IDH) mutation status is unequivocally established as the foremost molecular classifier, exerting decisive influence on diagnostic stratification, prognostication, and individualized treatment selection [2]. IDH encodes a metabolic enzyme indispensable for cellular energy homeostasis, and the near-universal presence of its mutations throughout the tumor together with their enduring retention across disease progression highlights the central role of these alterations in gliomagenesis [3]. The therapeutic benefit of both maximal safe resection and the emerging targeted agent vorasidenib is intrinsically contingent upon IDH mutational status [4–7]. Currently, definitive IDH genotyping still necessitates invasive tissue sampling, underscoring the pressing unmet need for non-invasive surrogates. Therefore, accurate, imaging-based preoperative ascertainment of IDH genotyping is pivotal to optimizing patient management and sharpening prognostic precision.

Diffusion magnetic resonance imaging techniques, such as diffusion weighted imaging (DWI) and diffusion tensor imaging (DTI), have been commonly used in the assessment of glioma characteristics at the cellular level, including tumor grade, Ki-67 expression, and IDH mutation status [8–10]. However, these methods are constrained by the Gaussian spin displacement assumption, limiting the precise description of microstructural alterations in complex brain tissue. Neurite orientation dispersion and density imaging (NODDI) has emerged as a solution that uses a multi-compartment diffusion model to separate signals from three tissue compartments (intra-neurite water, extra-neurite water and cerebrospinal fluid) [11]. Animal studies have validated the capacity of NODDI to quantitatively measure cerebral tissue density, correlating well with histological findings [12]. Moreover, NODDI enables analysis of neurite density and orientation dispersion by providing indices of the intracellular volume fraction (ICVF) indicating axon density, the orientation dispersion index (ODI) revealing axon organization, and the isotropic volume fraction (ISOVF) reflecting the proportion of isotropic water compartments in the tissue [13]. Recently, the applications of NODDI in neuroscience and clinical medicine have expanded, demonstrating significant potential as a non-invasive imaging biomarker [14, 15].

Currently, the application of NODDI for IDH genotyping in gliomas remains underexplored. The majority of prior studies have predominantly focused on the tumor parenchyma [16–18], a limitation that precludes a comprehensive evaluation of the intrinsic heterogeneity of gliomas. Although several recent studies have begun to examine the peritumoral areas, their findings have been inconsistent. For example, Zhao et al. calculated the mean NODDI parameters in both the tumor parenchyma and peritumoral regions but did not identify significant differences among IDH genotypes [19]. In contrast, Gao et al. performed histogram analysis on the whole-tumor region, including peritumoral edema, and demonstrated that NODDI exhibits moderate diagnostic performance in identifying the IDH genotype of gliomas, comparable to that of DTI [20]. Additionally, the results from Zerweck et al. indicate that the area under the curve (AUC) for IDH genotyping using ODI is higher in the peritumoral region than in the tumor parenchyma [21]. The above findings indicate potential variations in the diagnostic performance of NODDI across tumor subregions, with histogram features potentially enhancing their performance.

The tumor habitat refers to the unique microenvironment created by cancer cells and their surroundings as they adapt to various survival pressures [22]. Recently, habitat subregion analysis, which relies on multiparametric MRI signal attributes, has been proposed to generate distinct microenvironmental subregions. This approach aims to further identify and quantify the internal tumor microenvironment characteristics [23]. Currently, habitat analysis in gliomas mainly relies on traditional MRI sequences [24–29], although advanced diffusion models like NODDI are rarely explored in this context.

To our knowledge, no studies have systematically compared the diagnostic utility of DTI and NODDI histogram features within tumor spatial habitat subregions for IDH genotyping. Therefore, this study aims to quantitatively compare the diagnostic performance of NODDI and DTI histogram features across the whole-tumor region, NODDI-based subregions, and DTI-based subregions for IDH genotyping. This comparison will explore methods to enhance diagnostic performance and advance our understanding of tumor heterogeneity.

Materials and methods

Study participants

The study adhered to the Declaration of Helsinki (2013 revision). Approval was obtained from the ethics committee board and informed consent was obtained from all participating patients. Adult patients with suspected gliomas were prospectively and randomly enrolled at our institution between December 2021 and January 2025, with MRI scans performed within two weeks before their surgery and antitumor therapy began.

The following exclusion criteria were applied: (a) incomplete imaging data or poor image quality, namely, obvious artifacts or head motion; (b) without surgery or biopsy; (c) with histologically confirmed non-glioma lesions; (d) failure of genetic detection due to insufficient tissue.

An overview of the case selection process is shown in Fig. S1. Clinical information, including age, gender, tumor subtype, tumor grade, and the Ki-67 labeling index (Ki-67 LI), was recorded for each participant. Based on experience, assuming the moderate effect (d = 0.5), the significance level is 0.05 and the efficacy is equal to 0.8, the minimum sample size of each group is 32.

MRI parameters

Participants underwent diffusion-weighted imaging and structural MRI on a 3.0 T scanner (Magnetom Prisma; Siemens Healthcare) with a 64-channel head-neck coil. The structural MR imaging protocols included sagittal T1-weighted magnetization prepared rapid gradient echo sequence (T1-MPRAGE), axial T2-weighted (T2W) fast spin-echo (FSE) images, axial fluid-attenuated inversion recovery (FLAIR) T2W images, and contrast-enhanced axial/sagittal/coronal T1W (CE-T1W) images.

For diffusion-weighted imaging, a q-space sampling scheme was adopted for acquiring 128 diffusion samples, which consisted of 14 b-values (250, 500, 750, 1000, 1250, 1500, 2000, 2250, 2500, 2750, 3000, 3250, 3500, and 4000 s/mm²), along 3, 6, 4, 3, 12, 12, 6, 15, 12, 12, 4, 12, 24, and 3 directions, respectively. This q-space sampling scheme has emerged as a popular alternative alongside conventional multi-shell acquisitions and remains fully compatible with DTI and NODDI. The other scan parameters were: repetition time, 3900 ms; echo time, 88 ms; field of view, 230 × 230 mm²; GeneRalized Autocalibrating Partial Parallel Acquisition (GRAPPA), 2; slice acceleration factor, 2; number of averages, 1; voxel size, 2.5 × 2.5 × 2.5 mm³, without gap; pulse, bipolar; acquisition time, 8 min 44 s. The bipolar gradient scheme is employed to correct the distortions caused by eddy currents. For distortion correction, two b = 0 s/mm² images were acquired with opposite phase-encoding directions: anterior-to-posterior (AP) and posterior-to-anterior (PA).

Image processing

First, all diffusion-weighted images were converted from the DICOM format to the BIDS format. Subsequently, these images underwent a series of preprocessing steps using QSIprep (https://github.com/PennLINC/qsiprep), which included denoising, motion correction, and distortion correction. Specifically, distortion correction was performed with TOPUP by utilizing the b0 image pair acquired with opposite phase-encoding directions (AP and PA). During eddy-based head-motion and eddy-current correction, the corresponding b-vectors were automatically re-oriented; the resulting motion- and distortion-corrected diffusion-weighted images were then used for diffusion model estimation. DTI analysis was performed using b-values ranging from 200 to 1500 s/mm², and NODDI analysis utilized data with b-values not more than 3000 s/mm². The DTI-derived metrics, including fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD), were computed using FSL (https://fsl.fmrib.ox.ac.uk/fsl). The NODDI-derived metrics, including intracellular volume fraction (ICVF), orientation dispersion index (ODI), and isotropic volume fraction (ISOVF), were calculated using AMICO (https://github.com/daducci/AMICO).

Rigid registration was performed to align all structural MR images to the b0 images of the diffusion-weighted images using SPM12 (http://www.fil.ion.ucl.ac.uk/spm). The co-registered structural MR images served as structural references for volume of interest (VOI) delineation.

VOI placement and habitat analysis

Two radiologists, with 15 and 10 years of experience in neuroradiology respectively, manually delineated the VOIs covering the entire tumor region and peritumoral edema on T2W and FLAIR images in all participants by consensus, both blinded to clinical and pathological data. An additional 64 mm² square ROI was placed within the contralateral normal-appearing white matter (NAWM) in each patient.

Building upon previous studies that have highlighted the diagnostic performance of ICVF and ODI in NODDI for glioma grading and IDH-genotype prediction [17–20], ICVF and ODI maps were adopted as the primary quantitative parameters for habitat segmentation. In addition, FA and MD from DTI were employed to provide a comparative analysis with NODDI.

The Otsu threshold method was used to partition the voxels into high- and low-intensity regions for each map, with the threshold derived from the grayscale distribution of all voxels in VOIs across the entire cohort. The threshold value was determined iteratively, satisfying two critical conditions: (1) minimizing within-class variance and (2) maximizing between-class variance. Subsequently, the resulting subregions from each parameter set were amalgamated to yield four distinct final subregions. For the NODDI parameter set (ICVF and ODI), four regions were defined: Region 1 (ICVF_LODI_L) includes voxels with both ICVF and ODI below their thresholds; Region 2 (ICVF_LODI_H), ICVF below and ODI above the threshold; Region 3 (ICVF_HODI_L), ICVF above and ODI below the threshold; Region 4 (ICVF_HODI_H), both ICVF and ODI above their thresholds. Similarly, for the DTI parameter set (FA and MD), Region 1 (FA_LMD_L) includes voxels with both FA and MD below their thresholds; Region 2 (FA_LMD_H), FA below and MD above the threshold; Region 3 (FA_HMD_L), FA above and MD below the threshold; Region 4 (FA_HMD_H), both FA and MD above their thresholds. To assess the impact of the habitat classification method, we additionally applied K-means clustering. The algorithm partitioned the voxels in the whole-tumor region into four distinct clusters based on similarities and differences in ICVF and ODI distributions, with each cluster representing a habitat subregion.

Using the above methods, we obtained three distinct habitat segmentations: (1) Habitat D (D stands for DTI), segmented using the Otsu algorithm based on FA and MD from DTI; (2) Habitat N (N stands for NODDI), segmented using the Otsu algorithm based on ICVF and ODI from NODDI; and (3) Habitat K (K stands for K-means), segmented using the K-means algorithm based on ICVF and ODI from NODDI. Each habitat comprises four subregions. The workflow is shown in Fig. 1.

Fig. 1.

Overview of the analysis workflow. Dual-axis bubble quadrant chart of habitat analysis segmentation by DTI and NODDI parameters, with bubble sizes representing subregion volume proportions and corresponding values indicated in the lower right. R1-4 = Region 1–4; IDH = isocitrate dehydrogenase; DTI = diffusion tensor imaging; FA = fractional anisotropy; MD = mean diffusivity; AD = axial diffusivity; RD = radial diffusivity; NODDI = neurite orientation dispersion and density imaging; ICVF = intracellular volume fraction; ODI = orientation dispersion index; ISOVF = isotropic volume fraction

Histogram analysis

Histogram features were extracted by FeAture Explorer (FAE, Version 0.5.12) [30]. The following histogram features were extracted from the whole-tumor VOI and each habitat subregion: 10th percentile (P10), 90th percentile (P90), energy, entropy, interquartile range (IQR), kurtosis, maximum, mean absolute deviation (MAD), mean, median, minimum, range, robust mean absolute deviation (rMAD), root mean squared (RMS), skewness, total energy, uniformity and variance. For detailed calculation formulas, refer to the PyRadiomics documentation (available at https://pyradiomics.readthedocs.io/en/latest/features.html#module-radiomics.firstorder/). Definitions of each feature are provided in Table S1 in the supplementary material.

Molecular studies

Multiplexed polymerase chain reaction followed by next-generation sequencing was performed to reveal key molecular features. Single-nucleotide variant detection was conducted to detect mutations in IDH1 codon 132, IDH2 codon 172, TERT codon C228T, and TERT codon C250T. Copy number variant detection was used to identify heterozygous deletion on chromosome 1p/19q, loss of chromosome 10, and gain of chromosome 7. Real-time quantitative polymerase chain reaction was used to assess EGFRvIII amplification. A mutation in IDH1 codon 132 or IDH2 codon 172 supported a diagnosis of IDH-mutant glioma; otherwise, the glioma was considered IDH-wildtype. Patients were first grouped according to IDH genotype (IDH-wildtype or IDH-mutant). Subsequently, patients were placed in three subgroups according to current WHO CNS5 subtype criteria [2]. Additionally, the Ki-67 LI, a crucial marker of cellular proliferation, was evaluated using immunohistochemistry.

Statistical analysis

Statistical analyses were performed using IBM SPSS Statistics (version 26) and R (version 4.2.0), with p < 0.05 indicating significance. The Mann-Whitney U test was used to compare histogram features between IDH genotypes, followed by Benjamini-Hochberg correction for multiple comparisons. For logistic regression models predicting IDH genotype and grade, we selected histogram features with p < 0.05. The dataset was balanced using Synthetic Minority Over-sampling Technique (SMOTE), and features were standardized to a zero mean and unit variance. Highly correlated features (Pearson correlation coefficient, PCC > 0.8) were removed, and features with a variance inflation factor (VIF) > 5 were eliminated after assessing collinearity. Informative features were identified via Recursive Feature Elimination (RFE), and models were optimized through 5-fold cross-validation. Receiver Operating Characteristic (ROC) curve analyses were used to assess diagnostic performance, with the DeLong test employed to compare AUCs. A nomogram integrating selected features and clinical indicators was constructed for predictive modeling, and model performance was calibrated using calibration curves. The correlation between the Ki-67 LI and histogram features was quantified using the Pearson correlation coefficient.

Results

Participant demographics

A total of 75 participants (38 females, 37 males; mean age, 48 ± 13 years; age range, 20–72 years) were included in the study. Participant characteristics are detailed in Table 1. Significant differences were observed in age and Ki-67 LI between the IDH-wildtype and IDH-mutant groups (both p < 0.001). Patients with IDH-wildtype gliomas were older and had significantly higher Ki-67 LI values compared with those with IDH-mutant gliomas, consistent with previous reports [20, 31–34].

Table 1.

Demographic and clinical characteristics of participants

	IDH-wildtype	IDH-mutant	p
N	39	36	NA
Age	54 ± 12	42 ± 11	< 0.001*
Gender (Male/Female)	20/19	17/19	0.819
Tumor subtypes
Glioblastoma	34	NA	NA
Astrocytoma (grade 2/3/4)	NA	15/4/8
Oligodendroglioma (grade 2/3)	NA	5/4
Ki-67 LI	35 ± 23	17 ± 16	< 0.001*

IDH = isocitrate dehydrogenase; Ki-67 LI = Ki-67 labeling index; NA = Not Applicable. *represented a statistical difference (p < 0.05)

Habitats segmentation thresholds and subregion-specific characteristics

For Habitat D (DTI-based, Otsu-segmented), thresholds were set at FA = 0.23 and MD = 1.00 × 10⁻³ mm²/s. For Habitat N (NODDI-based, Otsu-segmented), cut-offs were applied at ICVF = 0.30 and ODI = 0.51. Volume fractions for each subregion are illustrated in Fig. 1. A scatter plot of subject-level means is shown in Fig. 2 to illustrate the characteristics of the habitat subregions; the corresponding spatial distribution on MRI is shown in Fig. 3. Specific mean values for FA, MD, ICVF, and ODI in NAWM and each subregion are listed in Supplementary Table S2.

Fig. 2.

Scatter plot of subject-level means. The solid white line indicates the mean of the contralateral normal-appearing white matter (NAWM), and the dashed line represents the segmentation threshold. Each circle denotes the mean value for an individual subject. MD units: 10⁻³ mm²/s. All directional arrows (↑ mild increase, ↑↑ marked increase, ↓ mild decrease, ↓↓ marked decrease) denote changes relative to the NAWM

Fig. 3.

Distribution of habitat subregions on MRI Images. Blue, orange, green, and red correspond to Regions 1–4, respectively

Comparative analysis of mean values and histogram features between IDH-wildtype and IDH-mutant groups

In the whole-tumor region, the mean FA and ICVF were higher in IDH-wildtype than IDH-mutant tumors (p < 0.001, p = 0.039, respectively). The mean MD was lower in IDH-wildtype but not significant after correction (p = 0.059). The mean ODI showed no significant difference. In the habitat subregions, the ODI of IDH-wildtype is lower than that of IDH-mutant group in Region 1 of Habitats N and Habitats K (both p < 0.001), but higher in Region 4 of Habitats N (p = 0.004). The specific intergroup differences of the mean values of each parameter in different regions are shown in Supplementary Fig. S2 and Table S3.

Histogram metrics were more sensitive than mean values. For instance, IDH-wildtype tumors showed lower P10 and median of ODI in the whole-tumor region (p = 0.021 and 0.012, respectively), lower minimum and P10 of FA in Region 4 of Habitat N (p = 0.005 and 0.025), and higher variance and MAD of ODI in Region 2 of Habitat K (both p = 0.001), while the corresponding mean values did not differ significantly.

Comparisons of the AUCs in identifying IDH genotypes

The optimal histogram features for the whole-tumor region and each habitat, along with their corresponding AUC, cut-off values, sensitivity, and specificity, are provided in Table 2. Among the DTI histogram features, the best AUC in the whole-tumor region was 0.787, while the highest AUC in habitats reached 0.864 (FA_P10 in Region 1 of Habitats N), a statistically significant improvement (DeLong p = 0.048). For NODDI features, the best AUC in the whole-tumor region was 0.726, compared to the highest AUC of 0.866 in habitats (ODI_median in Region 1 of Habitats K), which was also statistically significant (DeLong p = 0.004).

Table 2.

Performance of optimal diffusion MRI metrics in predicting IDH genotype across the whole tumor region and the habitat subregions

Habitat	Region	Feature	AUC (95%CI)	Cut-off	Sensitivity	Specificity
Whole-tumor region		DTI FA Median	0.787 (0.678, 0.873)	0.168	0.778	0.769
Habitats D	Region 2	DTI FA P90	0.752 (0.639, 0.844)	0.188	0.833	0.692
Habitats N	Region 1	DTI FA P10	0.864 (0.765, 0.932)	0.112	0.833	0.795
Habitats K	Region 1	DTI FA P10	0.838 (0.734, 0.916)	0.118	0.778	0.821
Whole-tumor region		NODDI ODI P10	0.726 (0.611, 0.823)	0.143	0.944	0.487
Habitats D	Region 2	NODDI ODI P10	0.722 (0.606, 0.819)	0.225	0.75	0.692
Habitats N	Region 1	NODDI ODI P90	0.801 (0.692, 0.884)	0.446	0.833	0.692
Habitats K	Region 1	NODDI ODI Median	0.866 (0.768, 0.934)	0.258	0.917	0.718

Values correspond to the cut-off determined by the Youden index. IDH = isocitrate dehydrogenase; AUC = area under the receiver operating characteristic curve; DTI = diffusion tensor imaging; FA = fractional anisotropy; NODDI = neurite orientation dispersion and density imaging; ODI = orientation dispersion index; P10 = 10th percentile; P90 = 90th percentile; Data in parentheses represent 95% confidence intervals (CIs)

A total of eight logistic regression models were constructed using histogram features derived from DTI and NODDI in the whole-tumor region (wr), Habitat D, Habitat N, and Habitat K. These models include four DTI models (DTI model wr, DTI model D, DTI model N, DTI model K) and four NODDI models (NODDI model wr, NODDI model D, NODDI model N, NODDI model K). The specific feature selection and diagnostic performance of each model are reported in detail in Table 3 and the ROC curves are shown in Supplementary Fig. S3.

Table 3.

Logistic regression models in the genotyping of IDH in gliomas

Model	β	p	AUC	Accuracy	Sensitivity	Specificity
DTI model wr			0.788 (0.679, 0.874)	0.760	0.806	0.718
Intercept	4.386	< 0.001
FA Median	-26.814	< 0.001
DTI model D			0.859 (0.759, 0.929)	0.760	0.778	0.744
Intercept	-6.768	0.330
R1 FA Variance	-1656.999	0.001
R1 RD P90	18.208	0.019
R2 FA RMS	-41.466	0.002
DTI model N			0.863 (0.764, 0.932)	0.800	0.833	0.769
Intercept	7.906	< 0.001
R1 FA P10	-72.858	< 0.001
DTI model K			0.901(0.810, 0.958)	0.853	0.872	0.833
Intercept	3.339	0.601
R1 FA P10	-58.558	0.008
R1 FA Minimum	-42.416	0.015
R2 FA Minimum	131.507	0.039
R2 FA Skewness	-0.050	0.957
R2 MD Entropy	1.147	0.425
R3 AD rMAD	-8.458	0.274
NODDI model wr			0.858 (0.758, 0.928)	0.787	0.833	0.744
Intercept	-0.526	0.714
ICVF Mean	-20.973	< 0.001
ODI P10	30.352	< 0.001
NODDI model D			0.886 (0.792, 0.948)	0.787	0.722	0.846
Intercept	5.009	0.062
R1 ICVF P10	-18.555	0.233
R1 ICVF Kurtosis	0.088	0.025
R1 ODI Variance	-80.584	0.038
R2 ISOVF rMAD	16.829	0.001
R3 ISOVF Maximum	-4.873	0.018
NODDI model N			0.944 (0.866, 0.984)	0.880	0.833	0.923
Intercept	20.668	0.285
R1 ICVF P10	-27.683	0.044
R1 ODI P90	59.009	0.002
R1 ODI Entropy	-10.748	0.024
R2 ICVF RMS	-9.929	0.290
R4 ODI Variance	-305.310	0.008
R4 ISOVF Variance	29.893	0.023
NODDI model K			0.927 (0.843, 0.975)	0.827	0.846	0.806
Intercept	6.645	0.116
R1 ICVF Minimum	-17.917	0.044
R1 ODI P10	21.829	0.101
R1 ODI Skewness	-4.533	0.042
R2 ICVF P10	6.063	0.670
R2 ODI Skewness	-5.002	0.008
R2 ODI Variance	-206.850	0.014
R3 ODI Minimum	-7.020	0.291

The model wr represents the logistic regression model constructed using the histogram features from the whole-tumor region. The models D/N/K represent logistic regression models constructed using histogram features from habitats D/N/K, respectively. IDH = isocitrate dehydrogenase. R1-4 = Region 1–4; DTI = diffusion tensor imaging; FA = fractional anisotropy; MD = mean diffusivity; NODDI = neurite orientation dispersion and density imaging; ICVF = intracellular volume fraction; ODI = orientation dispersion index

In model comparisons, the highest AUC for DTI models was 0.901 (Habitats K), significantly higher than the DTI model in whole-tumor region (AUC = 0.788, DeLong p = 0.028). For NODDI models, the highest AUC was 0.944 (Habitats N), significantly different from both the NODDI model in whole-tumor region (AUC = 0.858, DeLong p = 0.033) and the corresponding DTI model in Habitats N (AUC = 0.863, DeLong p = 0.034). More DeLong p-values are detailed in Supplementary Table S4. After cross-validation, the differences between the DTI models in the whole-tumor region and in Habitats N remained statistically significant (DeLong p = 0.036).

The AUC values for the optimal histogram features and each model, along with the corresponding DeLong test results, are shown in Fig. 4.

Fig. 4.

Bar Charts of AUC Values for IDH Genotyping. Panel A compares optimal histogram features in the whole-tumor region and habitats. Panel B compares model AUCs. Panel C shows 5-fold cross-validation AUCs. * indicates DeLong test p < 0.05. DTI = diffusion tensor imaging; FA = fractional anisotropy; MD = mean diffusivity; NODDI = neurite orientation dispersion and density imaging; ICVF = intracellular volume fraction; ODI = orientation dispersion index

Nomogram

The optimal logistic regression model for predicting IDH mutation status (NODDI model N), defined as the NODDI score, was used to construct a nomogram incorporating age, achieving an AUC of 0.962 (95% CI: 0.891–0.993). Calibration analysis showed good agreement between predicted probabilities and actual outcomes (Mean Absolute Error = 0.019). The Hosmer-Lemeshow test (p = 0.912) and Spiegelhalter Z-test (Z = 0.025, p = 0.980) both indicated excellent model calibration. The nomogram is shown in Fig. 5, with corresponding ROC and calibration curves presented in Supplementary Fig. S4. The NODDI score calculation formula is in Supplementary Equation S1.

Fig. 5.

Nomogram for IDH Genotyping. NODDI = neurite orientation dispersion and density imaging

Correlation analysis

In the whole-tumor region, the DTI feature most correlated with Ki-67 LI was FA_minimum (r = -0.456), while for NODDI it was ICVF_energy (r = 0.424). In the subregions, the highest DTI correlation was FA_minimum in Region 4 of Habitat N (r = -0.560), and for NODDI, it was ODI_MAD in Region 4 of Habitat N (r = 0.609). Scatter plots of these top-correlated features with Ki-67 LI are shown in Supplementary Fig. S5.

Exploratory analysis for predicting grades in IDH-mutant gliomas

In the whole-tumor region, the DTI model achieved an AUC of 0.862, while the NODDI model reached 0.950. Among the subregions, the DTI model in Habitats N performed the best with an AUC of 0.968 and was the only model significantly outperforming the whole-tumor DTI model. The corresponding ROC curve, AUC value and DeLong test results are shown in Supplementary Fig. S3 and.Fig. S6.

Discussion

In this study, we delineated habitat subregions in gliomas based on DTI and NODDI, and assessed histogram features differences across IDH genotypes. (1) In the whole-tumor region, NODDI and DTI models showed comparable performance in predicting IDH genotypes (AUC = 0.858 and 0.788, DeLong test p > 0.05). In Habitat D (DTI-based), neither model improved significantly (AUC = 0.886 and 0.859, DeLong test p > 0.05 both). In Habitat N (NODDI-based), both models improved, with NODDI outperforming DTI (AUC = 0.944 and 0.863, DeLong test p < 0.05). (2) Otsu and K-means differ in spatial segmentation but achieve consistent IDH diagnostic performance after feature extraction. (3) The nomogram integrating age and the optimal NODDI model achieved high diagnostic performance (AUC = 0.962). (4) DTI and NODDI features correlate better with Ki-67 LI in habitat subregions than in whole-tumor region. (5) Habitat analysis demonstrated potential in distinguishing pathological grades in IDH-mutant gliomas. In conclusion, habitat analysis enhances NODDI features for IDH genotype prediction, offering a promising tool for assessing proliferative activity and grading in gliomas.

NODDI, based on a multi-compartment model, depicts more microstructural changes than DTI. However, in whole-tumor region, histogram features showed comparable performance between the two models in predicting IDH genotypes, aligning with prior studies [20]. Further habitat analysis in our study revealed that in Habitats N, the diagnostic performance of models significantly improved, with NODDI outperforming DTI. In contrast, in Habitat D, neither model showed significant improvement, underscoring the enhanced capacity of NODDI to resolve tumor microenvironment heterogeneity.

We found that in Region 4 of Habitat N, IDH-wildtype tumors had a significantly higher ODI than IDH-mutant tumors. This region, characterized by mild reductions or increases in ICVF and significantly elevated ODI, indicates a highly disrupted tumor solid area. MRI images confirmed that this region corresponds to the core area of tumor parenchyma. Within this region, IDH-wildtype tumors also exhibited higher ODI dispersion metrics (IQR and MAD), suggesting more severe and heterogeneous disruption of white matter fiber bundles, consistent with previous studies [17, 19, 21].

However, in Region 1 of both Habitat N and Habitat K, IDH-wildtype tumors exhibited lower ODI values than IDH-mutant tumors, along with lower MD and higher FA. This region primarily corresponds to the edema area with mild disruption at the tumor margin, suggesting that IDH-wildtype tumors may preserve more intact white matter fiber bundles in the peritumoral edema. This contradictory findings in the peritumoral and solid regions align with Zerweck et al.‘s results [21]. Zhao et al. observed a similar trend, but it was not statistically significant, likely due to their small sample size and small ROI [19].This contrasting result between the peritumoral area and the solid area aligns with the findings of of Zerweck et al. and Zhao et al. [19, 21],. ODI has been shown to provide more stable and reliable white matter microstructure measurements than FA in the presence of edema [35], which may explain why NODDI-based habitats outperform DTI-based habitat. Additionally, in our study, IDH-wildtype tumors had significantly higher ICVF than IDH-mutant tumors in Region 2 of Habitat K (representing the tumor solid area). This aligns with previous literature [18–21] and may be due to higher density of palisading glioma cells.

Otsu and K-means are widely used image segmentation algorithms in tumor habitat analysis, each with pros and cons [32, 36–39]. Otsu thresholding excels in computational efficiency and interpretability, automatically optimizing grayscale intensity cutoffs, but its single-threshold limitation may compromise performance in complex tumor microenvironments with multimodal parameter distributions. Conversely, K-means clustering adaptively identifies intrinsic data patterns through iterative centroid optimization, though it is sensitive to initialization and susceptible to outliers. In our study, despite their differing spatial segmentation strategies, Otsu and K-means showed consistent IDH classification performance after feature extraction, highlighting the robustness of habitat segmentation across algorithms. Compared to manual delineation of solid regions and other ROIs slice by slice, habitat analysis is more simple and reduces human bias. Even though algorithm-generated subregions don not perfectly match traditional tissue components, the Otsu method combined with sequential parameter features can still objectively explain each subregion’s biological properties. This approach offers new potential for assessing tumor heterogeneity.

The optimal NODDI model was set as the NODDI score and combined with age to yield a high-precision nomogram for IDH genotyping (AUC = 0.962, 95% CI: 0.891–0.993), providing a clinically actionable tool for preoperative molecular stratification. Moreover, the correlation between habitat subregion histogram features and Ki-67 LI exceeds that of the whole-tumor region, with the highest correlation coefficients found in Region 4 of Habitat N, which corresponds to the high-cellularity solid tumor area. This finding may aid in guiding future biopsies to target the most aggressive regions.

While IDH-mutant gliomas have a survival advantage over IDH-wildtype gliomas, outcome disparities still exist between different grades within IDH-mutant gliomas [40]. Our study made an initial attempt to distinguish high and low grades within IDH-mutant gliomas to improve prognostic accuracy. NODDI-based habitat analysis enhanced DTI histogram features for grading IDH-mutant gliomas (AUC = 0.968 vs. 0.862, p < 0.05) but did not significantly enhance NODDI features.

Our study has several limitations. First, as a single-center study with a small sample size, our findings need validation in larger, multi-center cohorts. Cross-validation helps reduce overfitting but lacks the generalizability of an independent test set, which could lead to overfitting or underestimation of data variability. Future work should involve multi-center collaboration and include larger, independent datasets for robust model assessment. Second, habitat analysis lacks direct histological validation. While NODDI’s key parameters correlate with cell density and neurite orientation in prior studies, specific habitat subregions may need confirmation via stereotactic biopsy or spatial multi-omics technologies. Third, due to sample size constraints, we did not analyze 1p/19q codeletion status, which limits the comprehensive assessment of tumor heterogeneity. Fourth, genetic testing samples were sourced from representative solid tumor regions. However, glioma heterogeneity may restrict these results’ applicability to the whole tumor, impacting study interpretations and clinical applications. Future studies should incorporate multi-regional sampling for a more holistic genetic assessment.

Conclusion

In conclusion, NODDI is of significant value in predicting the IDH genotype of gliomas. Across the whole-tumor region, the histogram features of DTI and NODDI demonstrate comparable performance in glioma IDH genotyping. Habitat analysis based on NODDI can further enhance its performance, representing a promising approach that aids in exploring the microstructural heterogeneity of gliomas.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

DTI

Diffusion tensor imaging

DWI

Diffusion weighted imaging

FA

Fractional anisotropy

Habitat D

DTI-based Habitat (Otsu)

Habitat N

NODDI-based Habitat (Otsu)

Habitat K

NODDI-based Habitat (K-means)

ICVF

Isotropic volume fraction

IDH

Isocitrate dehydrogenase

IQR

Interquartile range

ISOVF

Isotropic volume fraction

Ki-67 LI

Ki-67 labeling index

MAD

Mean absolute deviation

MD

Mean diffusivity

NAWM

Contralateral normal white matter

NODDI

Neurite orientation dispersion and density imaging

ODI

Orientation dispersion index

P10

10th percentile

P90

90th percentile

rMAD

Robust mean absolute deviation

RMS

Root mean squared

VOI

Volume of interest

WHO

World health organization

Author contributions

YXJ Wáng contributed to the conception and design of the work. RF Jiang and YJ Xue contributed to the conception and design of the work and substantively revised it. ZT Yang and RL Lin made substantial contributions to the conception of the work, as well as the acquisition, analysis, and interpretation of data, and they also drafted the work and substantively revised it. ZX Wu, Y Song, and G Yang contributed to the acquisition, analysis, and interpretation of data. All authors read and approved the final version of the manuscript.

Funding

This work was supported by grants from Joint Funds for the innovation of science and Technology, Fujian province (2021Y9055), and Guidance Project of Fujian Science and Technology Program (2022Y0024).

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study adhered to the Declaration of Helsinki. Ethical approval was granted by the Ethics Committee of Fujian Medical University Union Hospital, and all participants provided informed consent. All methods were carried out according to relevant guidelines and regulations.

Consent for publication

Not applicable.

Competing interests

Yang Song is an employee of Siemens Healthineers Ltd. Shanghai, China. Other authors declare no conflict of interest.