Cellularity habitat-based MRI radiomics for non-invasive grading and IDH mutation prediction in adult-type diffuse glioma

Abstract

Background

Magnetic resonance imaging (MRI) radiomics has shown promise in glioma grading and isocitrate dehydrogenase (IDH) mutation prediction, but traditional whole-tumor approaches overlook intratumoral heterogeneity, limiting diagnostic accuracy and interpretability. This study aims to explore cellularity habitat-based MRI radiomics for precise grading and IDH mutation status prediction in adult-type diffuse glioma (ADG).

Methods

A total of 625 ADG patients were retrospectively collected. Whole-tumor volumes of interest (VOIs) were delineated on four conventional MRI sequences (T1WI, T2WI, T2-FLAIR, and CE-T1WI) and segmented into three cellularity habitats using apparent diffusion coefficient (ADC)-based K-means clustering: H₁ (low ADC), H₂ (medium ADC), and H₃ (high ADC). Radiomic features were extracted from individual and combined habitats, and predictive models were developed using a disentangled-learning-based multi-sequence fusion network (DMSFN). Performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE).

Results

The optimal habitats for ADG grading (Grade 2 vs. Grade 3 + 4, Grade 2 + 3 vs. Grade 4) and IDH prediction were H_1 + 2, H_1 + 2, and H_2 + 3, respectively. Combining T1WI, CE-T1WI, and T2-FLAIR sequences yielded the highest AUCs of 0.9360, 0.9605, and 0.8721 in the training set, and 0.8070, 0.8236, and 0.8180 in the independent test set. Shapley Additive exPlanation (SHAP) analysis identified key radiomic features contributing to model predictions, with CE-T1WI features consistently demonstrating high discriminative power.

Conclusions

Integrating ADC-derived cellularity habitats with MRI radiomics significantly improves the accuracy and biological interpretability of ADG grading and IDH mutation status prediction, offering a robust, non-invasive approach for glioma characterization.

Trial registration

Retrospectively registered.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-026-02182-w.

Introduction

Glioma is the most common and malignant tumor of the central nervous system (CNS) [1]. The fifth edition of the World Health Organization Classification of Central Nervous System Tumors (WHO CNS5) categorizes gliomas into pediatric and adult types, further classifying them into diffuse and localized types based on growth and infiltration characteristics [2]. Adult-type diffuse glioma (ADG), graded from 2 to 4, exhibits distinct histological and molecular features. Grade 2 represents low-grade glioma (LGG), which is less malignant and more responsive to radiotherapy and chemotherapy, while grades 3 and 4, classified as high-grade glioma (HGG), are highly invasive with a poorer prognosis. Additionally, isocitrate dehydrogenase (IDH) gene status serves as a key molecular marker for ADG, with IDH-mutant (IDH-MT) glioma demonstrating greater treatment sensitivity and better outcomes [3]. Current methods for ADG grading and IDH mutation assessment rely primarily on invasive procedures, such as biopsy or surgical resection. These approaches can be challenging for patients with tumors in inaccessible locations or poor health conditions [4], underscoring the critical need for non-invasive methods to precisely evaluate ADG grade and IDH status preoperatively, facilitating better diagnosis, informed treatment decisions, and improved prognostic predictions.

Magnetic resonance imaging (MRI) radiomics, which leverages quantitative imaging features, has emerged as a powerful tool for glioma grading and IDH status prediction [5–7]. However, traditional radiomics approaches often analyze voxel-based measurements across the entire tumor, neglecting inherent biological heterogeneity, which limits both their biological interpretability and clinical applicability [8]. To better capture tumor heterogeneity, some studies manually delineated the tumor into subregions with distinct imaging characteristics [9–11]. However, this method is subject to inter-observer variability, limited in identifying microscopic pathological differences, and time-consuming, limiting its scalability in large-scale clinical applications. Moreover, the predictive value of individual subregions is rarely evaluated in a systematic manner.

In recent years, habitat imaging has emerged as a promising data-driven technique for characterizing tumor heterogeneity [12–14]. By applying clustering algorithms, it automatically segments the tumor into subregions with distinct pathological and physiological characteristics, offering a more precise representation of the tumor microenvironment [15]. While previous studies have explored habitat models utilizing advanced MRI sequences, such as diffusion kurtosis imaging (DKI) and dynamic susceptibility contrast-enhanced (DSC) MRI, for IDH status prediction [16–18], these methods typically rely on simple statistical analysis of imaging parameters and specialized imaging protocols, limiting both their representational capacity and generalizability in routine clinical practice.

To address these limitations, we propose a clinically practical habitat imaging framework based on apparent diffusion coefficient (ADC) derived from diffusion-weighted imaging (DWI), a physiological MRI sequence widely used in clinical workflows. As a quantitative biomarker, ADC reflects water molecule diffusion, providing valuable insights into cellularity and tissue architecture [19]. Our method requires only whole-tumor delineation followed by K-means clustering to automatically segment subregions with distinct cellularities (i.e., habitats), thereby avoiding the inherent limitations of manual segmentation. Radiomic features are then extracted from conventional MRI sequences within each habitat and their combinations to enable spatially resolved and biologically interpretable predictions of ADG grade and IDH mutation status. This ADC-based habitat radiomics framework offers a reproducible and interpretable approach for glioma characterization in clinical settings.

Materials and methods

Patient population

The retrospective MRI data were collected from The University of California San Francisco Preoperative Diffuse Glioma MRI (UCSF-PDGM) from 2015 to 2021, and an internal institutional dataset from Beijing Tiantan Hospital (BJTTH) in China from 2009 to 2021. The internal dataset was approved by the Institutional Ethics Committee of BJTTH (Approval No. KY2022-078-04), with a waiver of consent for the use of retrospective or public datasets.

The inclusion criteria were: (1) patients diagnosed with ADG according to the WHO CNS5 guideline, (2) patients with complete preoperative multiparametric MRI data, including T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), T2 fluid-attenuated inversion recovery (T2-FLAIR), contrast-enhanced T1-weighted imaging (CE-T1WI), and DWI. The patients with MR images of poor quality were excluded. A total of 625 ADG patients met the criteria, with 476 from UCSF-PDGM serving as the training set and 149 from the internal institution used as the independent test set.

Clinical characteristics for both the training and test sets are presented in Table 1a, 1b and 1c. All tumors were tested for IDH mutations through genetic sequencing of tissue obtained during biopsy or resection.

Table 1a.

Patient characteristics for Grade 2 vs. Grade 3 + 4

		Training set (n = 476)			Test set (n = 149)
		Grade 2	Grade 3 + 4	p-value	Grade 2	Grade 3 + 4	p-value
Total		46 (9.7%)	430 (90.3%)	/	44 (29.53%)	105 (70.47%)	/
Age		38.09 ± 11.83	58.81 ± 14.15	< 0.001a*	39.55 ± 11.33	50.54 ± 13.42	< 0.001a*
Gender
	0, Female	21 (4.4%)	170 (35.7%)	0.421b	16 (10.74%)	51 (34.23%)	0.172b
	1, Male	25 (5.3%)	260 (54.6%)		28 (18.79%)	54 (36.24%)

^*p < 0.05; ^a the Student’s t-test; ^b the Chi-square test

Table 1b.

Patient characteristics for Grade 2 + 3 vs. Grade 4

		Training set (n = 476)			Test set (n = 149)
		Grade 2 + 3	Grade 4	p-value	Grade 2 + 3	Grade 4	p-value
Total		75 (15.8%)	401 (84.2%)	/	74 (49.7%)	75 (50.3%)	/
Age		39.12 ± 11.51	60.12 ± 13.45	< 0.001a*	40.82 ± 11.50	53.58 ± 12.83	< 0.001a*
Gender
	0, Female	32 (6.7%)	159 (33.4%)	0.625b	31 (20.8%)	36 (24.2%)	0.454b
	1, Male	43 (9.1%)	242 (50.8%)		43 (28.9%)	39 (26.1%)

^*p < 0.05; ^a the Student’s t-test; ^b the Chi-square test

Table 1c.

Patient characteristics for IDH-WT vs. IDH-MT

		Training set (n = 476)			Test set (n = 149)
		IDH-WT	IDH-MT	p-value	IDH-WT	IDH-MT	p-value
Total		373 (78.4%)	103 (21.6%)	/	68 (45.6%)	81 (54.4%)	/
Age		61.78 ± 12.04	38.81 ± 11.52	< 0.001a*	55.18 ± 11.77	40.68 ± 11.70	< 0.001a*
Gender
	0, Female	151 (31.7%)	40 (8.4%)	0.763b	33 (22.1%)	34 (22.8%)	0.423b
	1, Male	222 (46.7%)	63 (13.2%)		35 (23.5%)	47 (31.6%)

^*p < 0.05; ^a the Student’s t-test; ^b the Chi-square test. IDH, isocitrate dehydrogenase; IDH-WT, IDH-wildtype; IDH-MT, IDH-mutant.

Image acquisition

All patients underwent imaging using either a 1.5 T or 3.0 T clinical MR scanner. The 1.5 T scanner was a GE Healthcare Explorer (Paris, France). The 3.0 T scanners included Siemens Healthcare Verio/Prisma (Erlangen, Germany), GE Healthcare Premier/Architect/Discovery MR 750 (Milwaukee, WI, USA), and Philips Healthcare Ingenia CX (Best, Netherlands). The brain imaging protocol included the following sequences: (a) T1WI, (b) T2WI, (c) T2-FLAIR, (d) CE-T1WI, and (e) DWI. Two gadolinium-based contrast agents were used: gadobutrol (Gadovist; Bayer) at a dose of 0.1 mL per kilogram of body weight and gadoterate (Dotarem, Guerbet) at a dose of 0.2 mL per kilogram of body weight. Detailed scanning parameters are provided in Table S1.

Image preprocessing and segmentation

For the UCSF-PDGM dataset, the image preprocessing and segmentation procedures are available on the The Cancer Imaging Archive (TCIA) website (https://www.cancerimagingarchive.net/collection/ucsf-pdgm). For the internal dataset, images underwent the same standardized preprocessing pipeline, including N4 bias field correction, skull stripping using HD-BET, non-linear registration of all sequences to the T2-FLAIR sequence using Advanced Normalization Tools (ANTs), resampling to a uniform voxel resolution of 1 × 1 × 1 mm³, and z-score intensity normalization. The whole-tumor volume of interest (VOI) was delineated slice-by-slice on T2-FLAIR images using ITK-SNAP by two radiologists, each with 2 years of neuroimaging experience. The delineations were then validated by a senior neuroradiologist with 19 years of neuroimaging experience. The VOI was defined as high-signal regions within the white matter, including the enhancing tumor, non-enhancing tumor, and peritumoral edema.

Construction of cellularity habitats

The voxels within the VOI were categorized into distinct cellularity habitats on each ADC map using the K-means clustering, applied independently at the patient level. The optimal number of habitats (corresponding to the “n_clusters” parameter in K-means, denoted as k) was determined using the elbow method [20], applied to the aggregated voxels from all patients in the training set. Quantitative evaluation showed that the sum of squared errors (SSE) decreased markedly as k increased from 1 to 3, dropping from 4.34 × 10⁷ to 0.91 × 10⁷, with relative reductions of 63.18% and 43.28% for k = 2 and k = 3, respectively. For k values greater than 3, the marginal reduction in SSE became substantially smaller, suggesting a clear elbow point at k = 3. Balancing clustering performance and model complexity, k = 3 was selected as the optimal number of habitats (Fig. S1, Table S2). Accordingly, the tumor VOI of each patient was partitioned into three cellularity habitats: Habitat 1 (H₁), characterized by low ADC values; Habitat 2 (H₂), characterized by medium ADC values; and Habitat 3 (H₃), characterized by high ADC values. Representative examples of each habitat are shown in Fig. S2. Finally, a total of seven habitats were obtained, including the three individual habitats (H₁, H₂, and H₃), the unions of each pair (H_{1 + 2}, H_{1 + 3}, and H_{2 + 3}), and the union of all three habitats (H_{1 + 2 + 3}), which represents the entire tumor.

Radiomic feature extraction

A total of 94 radiomic features were extracted from each available MRI sequence within each habitat using Pyradiomics (https://pyradiomics.readthedocs.io/en/latest/). These features included: (1) first-order statistics features, and (2) texture features, which encompassed the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size-zone matrix (GLSZM), gray-level dependence matrix (GLDM), and neighboring gray-tone-difference matrix (NGTDM). The radiomic features extracted from T1WI, CE-T1WI, T2WI, and T2-FLAIR were denoted as F₁, F₂, F₃, and F₄, respectively. Prior to model training, all radiomic features underwent z-score normalization.

Model development and validation

A disentangled-learning-based multi-sequence fusion network (DMSFN) was developed to integrate multiple MRI sequences for ADG grading and IDH mutation status prediction (Fig. 1). In the DMSFN, separate network branches process each MRI sequence, with an encoder network designed to extract hidden representations that are disentangled into sequence-unique and sequence-shared components. During model training, all radiomic features were provided as inputs. Feature dimensionality was handled implicitly through progressive compression within the sequence-specific encoders, enabling learned low-dimensional latent representations rather than explicit pre hoc feature selection. These components are then processed through a decoder network to reconstruct the original feature values for each sequence. The sequence-unique and sequence-shared components from all MRI sequences are integrated to form a fused representation, which is input into a prediction network tailored for the downstream tasks of ADG grading or IDH status prediction. Further details about the architecture, training process, and parameter settings of the DMSFN model are provided in the Supplementary Appendix.

Fig. 1.

Flowchart of the study. H, Habitat; IDH, isocitrate dehydrogenase; IDH-WT, IDH-wildtype; IDH-MT, IDH-mutant

Model validation was performed using a 5-fold cross-validation approach, repeated five times for each combination of individual MRI sequences and habitats within the training set to identify the optimal habitat. Cross-validation was implemented using StratifiedKFold with shuffle enabled and a fixed random seed (random_state = 0). To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied within each cross-validation iteration. Specifically, SMOTE oversampled the minority class in the radiomic features of the training subset (4 folds) to match the majority class, while the validation subset (1 fold) remained unaltered to prevent data leakage. Using the selected optimal habitat, multiple sequences were analyzed in conjunction to determine the optimal sequence combination. Finally, the optimal sequence combination was validated on an independent test set, with results compared against single-sequence models. Test-set performance uncertainty was quantified using bootstrap resampling (1,000 iterations). For each iteration, the model was retrained on a bootstrap sample of the training set and evaluated on the fixed test set. Performance metrics are reported as mean (95% confidence interval, CI), including the area under the receiver operating characteristic (ROC) curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE).

Statistical analysis

Continuous variables were expressed as means ± standard deviations, while categorical variables were presented as counts (percentages). The Chi-square test was used for categorical variables. For continuous variables, the Student’s t-test was applied to normally distributed data, and the Mann-Whitney U test was employed for non-normally distributed data. Comparisons of AUCs between different sequence combinations were conducted using the DeLong test. Adjusted p-values by Benjamini–Hochberg false discovery rate (FDR) correction < 0.05 was considered statistically significant. Statistical analyses were performed using SPSS version 21.0 (SPSS, Inc.) and Python version 3.6.

Results

Patient characteristics

A total of 625 patients were included, with 476 from UCSF-PDGM serving as the training set and 149 from BJTTH as the independent test set. Table 1a, 1b and 1c provide a comprehensive overview of the patient characteristics for both the training and test sets. No significant difference in gender distribution was observed between groups in both the training and test sets (p > 0.05 for all tasks). However, a significant difference in age was noted (p < 0.001 for all tasks).

Optimal tumor habitat and MRI sequence combination

Radiomic features from seven cellularity habitats corresponding to each conventional MRI sequence were extracted and subsequently utilized for separate training within DMSFN. The mean results from 5-fold cross-validation are presented in Table S3 and Fig. S3.

Cellularity habitat-based radiomics outperformed whole-tumor analysis in both grading and IDH mutation status prediction. For ADG grading (Grade 2 vs. Grade 3 + 4 and Grade 2 + 3 vs. Grade 4), Habitat 12 (H_1 + 2) yielded the highest average AUC values of 0.8678 and 0.8668, surpassing the whole-tumor AUCs of 0.8516 and 0.8574. For IDH mutation status prediction, Habitat 23 (H_2 + 3) achieved the highest average AUC of 0.8365, higher than the whole-tumor AUC of 0.8246. Therefore, H_1 + 2 will be employed for grading in subsequent experiments, while H_2 + 3 will be used for IDH mutation prediction.

In the analysis of the four MRI single sequences, the CE-T1WI sequence exhibited the highest performance across both ADG grading and IDH mutation prediction, with average AUCs of 0.8810, 0.9108, and 0.8595, respectively. This was followed by the T2-FLAIR sequence, with average AUCs of 0.8402, 0.8810, and 0.8456. The lowest performances were observed with average AUCs of 0.8376, 0.8244, and 0.7767 for T2WI, and 0.8130, 0.7555, and 0.7459 for T1WI, respectively.

Additionally, cellularity habitat-based radiomic features extracted from the ADC sequence further demonstrated superior performance over whole-tumor-based analysis in both ADG grading and IDH mutation prediction, as summarized in Table S4. The performance of different MRI sequence combinations based on the optimal habitat is provided in Table S5 and Fig. 2. The optimal combination identified through model validation for all three tasks consisted of T1WI, CE-T1WI, and T2-FLAIR, achieving AUCs of 0.9360, 0.9605, and 0.8721 in the training set, significantly outperforming both individual sequences and most other sequence combinations (Table 2 and Table S6).

Fig. 2.

AUCs of different MRI sequence combinations based on the optimal habitat. A AUCs of Grade 2 vs. Grade 3 + 4. B AUCs of Grade 2 + 3 vs. Grade 4. C AUCs of IDH-WT vs. IDH-MT. The asterisk indicates the optimal sequence combination. The numbers 1, 2, 3, and 4 represent the T1WI, CE-T1WI, T2WI, and T2-FLAIR sequences, respectively. AUC, area under the curve; F, feature; IDH, isocitrate dehydrogenase; IDH-WT, IDH-wildtype; IDH-MT, IDH-mutant

Table 2.

Performance of the optimal sequence combination on the training set

	Group	AUC	ACC	SEN	SPE	P
Grade 2 vs. Grade 3 + 4	F1 + H1 + 2	0.8263 ± 0.0429	0.7369 ± 0.0943	0.7215 ± 0.1143	0.8548 ± 0.0680	0.0045*
	F2 + H1 + 2	0.8911 ± 0.0495	0.8304 ± 0.0382	0.8223 ± 0.0507	0.8936 ± 0.0917	0.0314*
	F4 + H1 + 2	0.8908 ± 0.0162	0.8477 ± 0.0718	0.8459 ± 0.0921	0.8642 ± 0.1108	0.0339*
	F1 + 2 + 4+H1 + 2	0.9360 ± 0.0329	0.8704 ± 0.0515	0.8598 ± 0.0610	0.9527 ± 0.0686	/
Grade 2 + 3 vs. Grade 4	F1 + H1 + 2	0.7729 ± 0.0601	0.7559 ± 0.0717	0.7709 ± 0.1058	0.6952 ± 0.0858	< 0.0001*
	F2 + H1 + 2	0.9294 ± 0.0513	0.8899 ± 0.0714	0.8881 ± 0.0875	0.8976 ± 0.0342	0.0414*
	F4 + H1 + 2	0.9119 ± 0.0514	0.8536 ± 0.0228	0.8451 ± 0.0570	0.8865 ± 0.1611	0.0064*
	F1 + 2 + 4+H1 + 2	0.9605 ± 0.0208	0.9052 ± 0.0326	0.8920 ± 0.0351	0.9594 ± 0.0400	/
IDH-WT vs. IDH-MT	F1 + H2 + 3	0.8053 ± 0.0577	0.7461 ± 0.0445	0.7525 ± 0.1328	0.7440 ± 0.0735	0.0002*
	F2 + H2 + 3	0.8623 ± 0.1138	0.8199 ± 0.1200	0.7734 ± 0.1116	0.8318 ± 0.1464	0.2007
	F4 + H2 + 3	0.8558 ± 0.0214	0.8078 ± 0.0508	0.7487 ± 0.1135	0.8228 ± 0.0888	0.0632
	F1 + 2 + 4+H2 + 3	0.8721 ± 0.0824	0.8187 ± 0.0861	0.8098 ± 0.1192	0.8214 ± 0.1032	/

*FDR-adjusted p < 0.05. Data are presented as mean ± standard deviation. Bold values indicate the optimal performance. F, feature; H, habitat; IDH, isocitrate dehydrogenase; IDH-WT, IDH-wildtype; IDH-MT, IDH-mutant; AUC, area under the curve; ACC, accuracy; SEN, sensitivity; SPE, specificity; FDR, false discovery rate

Evaluation of the optimal sequence combination on the test set

Validation of the optimal habitat-based combination on the test set for ADG grading and IDH mutation status prediction is summarized in Table 3, with corresponding ROC curves presented in Fig. 3.

Table 3.

Performance of the optimal sequence combination on the test set

	Group	AUC	ACC	SEN	SPE	P
Grade 2 vs. Grade 3 + 4	F1 + H1 + 2	0.6211 (0.5803–0.6433)	0.6479 (0.5535–0.7315)	0.6916 (0.5902–0.7821)	0.5436 (0.4498–0.6273)	0.0027*
	F2 + H1 + 2	0.7501 (0.7097–0.7850)	0.6802 (0.5871–0.7552)	0.6156 (0.5286–0.6864)	0.8343 (0.6472–0.9773)	0.0492*
	F4 + H1 + 2	0.7052 (0.6812–0.7273)	0.7396 (0.6644–0.7852)	0.6695 (0.5091-0.7500)	0.8109 (0.6426–0.9143)	0.0205*
	F1 + 2 + 4+H1 + 2	0.8070 (0.7962–0.8271)	0.7642 (0.6945–0.8121)	0.7641 (0.6045–0.8571)	0.7643 (0.6591–0.9102)	/
Grade 2 + 3 vs. Grade 4	F1 + H1 + 2	0.6945 (0.6187–0.7817)	0.6732 (0.6107–0.7351)	0.6581 (0.5793–0.8337)	0.6885 (0.5389–0.7919)	0.0235*
	F2 + H1 + 2	0.7436 (0.7218–0.7821)	0.7070 (0.6779–0.7450)	0.6980 (0.5397–0.8533)	0.7161 (0.5534–0.8591)	0.0465*
	F4 + H1 + 2	0.7359 (0.7192–0.7632)	0.7114 (0.6846–0.7450)	0.7116 (0.5930–0.8337)	0.7160 (0.5804–0.8043)	0.0428*
	F1 + 2 + 4+H1 + 2	0.8236 (0.8078–0.8452)	0.7702 (0.7450–0.7919)	0.7609 (0.6267–0.8737)	0.7796 (0.6622–0.9125)	/
IDH-WT vs. IDH-MT	F1 + H2 + 3	0.7145 (0.6839–0.7776)	0.7014 (0.6711–0.7450)	0.7669 (0.5426–0.8830)	0.6234(0.5588–0.7460)	0.0340*
	F2 + H2 + 3	0.7598 (0.7303–0.7968)	0.7108 (0.6743–0.7450)	0.6940 (0.5309–0.8272)	0.7309 (0.6029–0.8529)	0.0987
	F4 + H2 + 3	0.6856 (0.6560–0.7269)	0.6577 (0.6242–0.6980)	0.6772 (0.5185–0.8583)	0.6346 (0.5188–0.7096)	0.0241*
	F1 + 2 + 4+H2 + 3	0.8180 (0.7847–0.8476)	0.7663 (0.7347–0.7919)	0.8093 (0.6849–0.9324)	0.7151 (0.6735–0.8235)	/

^*FDR-adjusted p < 0.05. Data are presented as mean (95% confidence interval). Bold values indicate the optimal performance. F, feature; H, habitat; IDH, isocitrate dehydrogenase; IDH-WT, IDH-wildtype; IDH-MT, IDH-mutant; AUC, area under the curve; ACC, accuracy; SEN, sensitivity; SPE, specificity; FDR, false discovery rate

Fig. 3.

ROC curves of the optimal sequence combination. A ROC curve of Grade 2 vs. Grade 3 + 4 on the training set. B ROC curve of Grade 2 + 3 vs. Grade 4 on the training set. C ROC curve of IDH-WT vs. IDH-MT on the training set. D ROC curve of Grade 2 vs. Grade 3 + 4 on the test set. E ROC curve of Grade 2 + 3 vs. Grade 4 on the test set. F ROC curve of IDH-WT vs. IDH-MT on the test set. ROC, receiver operating characteristic; IDH, isocitrate dehydrogenase; IDH-WT, IDH-wildtype; IDH-MT, IDH-mutant

For Grade 2 vs. Grade 3 + 4, the optimal combination yielded an AUC of 0.8070, with ACC, SEN, and SPE values of 0.7642, 0.7641, and 0.7643, respectively. For Grade 2 + 3 vs. Grade 4, it achieved an AUC of 0.8236, with ACC, SEN, and SPE values of 0.7702, 0.7609, and 0.7796, respectively. For IDH mutation status prediction, the optimal combination demonstrated an AUC of 0.8180, with corresponding ACC, SEN, and SPE values of 0.7663, 0.8093, and 0.7151.

The optimal combination consistently outperformed individual sequences, demonstrating statistically significant improvements in AUC (p < 0.05 for almost all tasks, except for F₂ [CE-T1WI] in IDH prediction).

Model interpretability and feature analysis

The Shapley Additive exPlanation (SHAP) method was employed to interpret the DMSFN model using the optimal multi-sequence combination. SHAP values were computed on the training set using the DeepExplainer module from the Python shap package, with the original multi-sequence radiomic features and corresponding model-predicted target-class probabilities as inputs. This enabled identification of the radiomic features most influential to the model’s predictions.

For Grade 2 vs. Grade 3 + 4, the top-5 contributing features included two first-order features and one texture feature from CE-T1WI, along with two texture features from T2-FLAIR (Fig. 4). For Grade 2 + 3 vs. Grade 4, all top-5 features were derived from CE-T1WI, comprising two first-order and three texture features. Similarly, for IDH mutation status prediction, the most significant features were exclusively from CE-T1WI, consisting of three first-order and two texture features.

Fig. 4.

SHAP violin plots. A SHAP values of Grade 2 vs. Grade 3 + 4. B SHAP values of Grade 2 + 3 vs. Grade 4. C SHAP values of IDH-WT vs. IDH-MT. SHAP, Shapley Additive exPlanation; IDH, isocitrate dehydrogenase; IDH-WT, IDH-wildtype; IDH-MT, IDH-mutant

Except for the “Minimum” first-order feature in the IDH-wildtype (IDH-WT) glioma vs. IDH-mutant (IDH-MT) glioma comparison, all other features demonstrated statistically significant differences across the groups (Grade 2 vs. Grade 3 + 4, Grade 2 + 3 vs. Grade 4, or IDH-WT vs. IDH-MT) (Table 4). Furthermore, except for the “Minimum” feature in Grade 2 vs. Grade 3 + 4 and IDH-WT vs. IDH-MT, the remaining features exhibited strong discriminative power, accurately distinguishing at least 60% of the patients.

Table 4.

Top-5 features contributing most to the model output of the optimal combination

Tasks	Category	Top-5 Features	Sequence	P-value	M	(< M | > M)
	First order	Skewness (1st)	CE-T1WI	< 0.001*	1.1206	Grade 2 (80.36% | 19.64%) Grade 3 + 4 (34.68% | 65.32%)
Grade 2	First order	Minimum (2nd)	CE-T1WI	0.002*	19.3154	Grade 2 (41.07% | 58.93%) Grade 3 + 4 (59.91% | 40.09%)
vs.	Texture	GLRLM_RunEntropy (3rd)	T2-FLAIR	< 0.001*	4.2217	Grade 2 (64.29% | 35.71%) Grade 3 + 4 (32.88% | 67.12%)
Grade 3 + 4	Texture	GLSZM_ZoneEntropy (4th)	T2-FLAIR	< 0.001*	5.0389	Grade 2 (62.50% | 35.50%) Grade 3 + 4 (34.01% | 65.99%)
	Texture	GLCM_DifferenceEntropy (5th)	CE-T1WI	< 0.001*	0.7329	Grade 2 (89.29% | 10.71%) Grade 3 + 4 (20.50% | 79.50%)
	First order	Skewness (1st)	CE-T1WI	< 0.001*	1.1746	Grade 2 + 3 (74.75% | 25.25%) Grade 4 (34.16% | 65.84%)
Grade 2 + 3	Texture	GLCM_MCC (2nd)	CE-T1WI	< 0.001*	0.6910	Grade 2 + 3 (70.71% | 29.29%) Grade 4 (17.46% | 82.54%)
vs.	Texture	GLDM_DependenceVariance (3rd)	CE-T1WI	< 0.001*	40.8556	Grade 2 + 3 (75.76% | 24.24%) Grade 4 (32.17% | 67.83%)
Grade 4	First order	StandardDeviation (4th)	CE-T1WI	< 0.001*	14.6645	Grade 2 + 3 (95.96% | 4.04%) Grade 4 (20.95% | 79.05%)
	Texture	GLSZM_ZoneEntropy (5th)	CE-T1WI	< 0.001*	4.4819	Grade 2 + 3 (85.86% | 14.14%) Grade 4 (16.96% | 83.04%)
	First order	Skewness (1st)	CE-T1WI	< 0.001*	1.6210	IDH-WT (37.53% | 62.47%) IDH-MT (65.05% | 34.95%)
IDH-WT	Texture	GLDM_DependenceVariance (2nd)	CE-T1WI	< 0.001*	44.3545	IDH-WT (27.71% | 72.29%) IDH-MT (82.52% | 17.48%)
vs.	First order	Minimum (3rd)	CE-T1WI	0.0643	16.1716	IDH-WT (49.62% | 50.38%) IDH-MT (59.22% | 40.78%)
IDH-MT	Texture	GLRLM_GrayLevelVariance (4th)	CE-T1WI	< 0.001*	0.7912	IDH-WT (32.49% | 67.51%) IDH-MT (93.20% | 6.80%)
	First order	StandardDeviation (5th)	CE-T1WI	< 0.001*	14.8812	IDH-WT (29.47% | 70.53%) IDH-MT (90.29% | 9.71%)

*p < 0.05 (unadjusted; FDR correction not applied as only a single pairwise comparison was performed). The p-value was calculated using the Mann–Whitney U test. The letter “M” represents the mean feature value of the positive group (Grade 3 + 4, Grade 4, or IDH-MT) and the mean feature value of the negative group (Grade 2, Grade 2 + 3, or IDH-WT). The notation “(< M | > M)” indicates the percentage of patients in the positive and negative groups with feature values below or above the “M” value. Bold values highlight discriminative features that can correctly distinguish at least 60% of patients in both the positive and negative groups. MCC, maximal correlation coefficient; IDH, isocitrate dehydrogenase; IDH-WT, IDH-wildtype; IDH-MT, IDH-mutant; GLCM, gray-level co-occurrence matrix; GLRLM, gray-level run-length matrix; GLSZM, gray-level size-zone matrix; GLDM, gray-level dependence matrix

The heatmap of the fused representation from the optimal combination further highlighted its robust discriminative capacity, as shown in Fig. S4.

Discussion

Accurate preoperative ADG grading and IDH mutation status prediction are crucial for tailoring treatment strategies and prognosis. While MRI radiomics techniques have been widely applied, existing studies often lack biological interpretability or rely on manual subregion segmentation, which is subjective, imprecise, and labor-intensive. To overcome these limitations, our study introduced an ADC-derived cellularity habitat-based MRI radiomics model. This model automatically segments tumor regions into distinct habitats using a data-driven K-means clustering algorithm. By extracting radiomic features from these biologically meaningful habitats, our model significantly improved the accuracy of ADG grading (Grade 2 vs. Grade 3 + 4, Grade 2 + 3 vs. Grade 4) and IDH mutation status prediction. The model achieved AUCs of 0.9360, 0.9605, and 0.8721 in the training set, and 0.8070, 0.8236, and 0.8180 in the independent test set.

Previous studies have explored the potential of habitat models using advanced MRI sequences for IDH prediction. For example, Liu et al. utilized DKI-derived habitat features to capture glioma spatial heterogeneity [16]. Álvarez-Torres et al. identified vascular habitat features from DSC-MRI to distinguish IDH-MT from IDH-WT glioma [17]. Wu et al. further validated the predictive power of DSC-MRI-based vascular habitat analysis for IDH status and prognosis in HGG [18]. However, these studies primarily relied on simple statistical descriptors of imaging parameters (e.g., maximum, minimum, and average values, or habitat volume percentages), which cannot comprehensively capture tumor heterogeneity. Furthermore, the reliance on advanced MRI sequences limits clinical applicability and generalizability. In contrast, our study integrates cellularity habitats derived from ADC, a widely available clinical MRI sequence, into radiomics analysis. ADC reflects tumor malignancy through cellularity. Our findings demonstrate superior predictive accuracy while maintaining broad clinical feasibility.

Traditional radiomics studies have predominantly focused on whole-tumor regions for glioma grading and IDH mutation status prediction, often overlooking the contributions of intratumoral heterogeneity. Some studies attempted to capture this heterogeneity by manually segmenting tumors into subregions (e.g., necrotic cores, enhancing tumors, and peritumoral edema) [9–11, 21]. However, these methods typically aggregate features from all subregions without evaluating the predictive value of each individually. By systematically exploring ADC-derived cellularity habitats and their combinations using the K-means clustering algorithm, our study identified the most predictive subregion, enhancing both predictive accuracy and biological interpretability.

For ADG grading, the most predictive habitat was H_{1 + 2}, encompassing regions with the lowest ADC (H₁) and moderate ADC (H₂) values. For IDH status prediction, H_{2 + 3} was most informative, covering areas with moderate ADC (H₂) and the highest ADC (H₃) values. Glioma grading is histopathologically characterized by increased cellularity, mitotic activity, and necrosis [2], which manifest as restricted diffusion on ADC maps [22]. The strong predictive performance of H_{1 + 2} likely reflects its ability to capture these highly cellular regions typical of higher-grade tumors. In contrast, IDH-mutant gliomas undergo metabolic reprogramming driven by 2-hydroxyglutarate (2-HG), which alters epigenetic regulation and cellular metabolism [23–25]. These tumors generally show slower growth kinetics and more preserved tissue structure [3], features that tend to present with less restricted diffusion, potentially explaining the predictive value of H_{2 + 3} in capturing these metabolic and microstructural differences. These results suggest that regions with higher cellularity may contribute more strongly to grading-related microstructural differences. In contrast, areas with relatively lower cellularity may provide complementary information relevant to IDH-mutant tumor biology.

Habitat-based radiomics consistently outperformed whole-tumor analysis across all tasks. For grading, the optimal habitat (H_{1 + 2}) achieved AUCs of 0.8678 and 0.8668, compared with 0.8516 and 0.8574 for whole-tumor analysis. For IDH prediction, the optimal habitat (H_{2 + 3}) achieved an AUC of 0.8365 versus 0.8246, consistent with previous reports [26, 27]. Although the absolute performance gains were modest, likely due to partial overlap with whole-tumor features and the coarse spatial partition (k = 3), habitat-based analysis offers enhanced interpretability through spatially localized feature attribution [18, 28].

Among single sequences, CE-T1WI emerged as the most predictive for both ADG grading and IDH mutation prediction, reflecting the strong association between tumor-enhancing regions and active proliferation [29, 30]. Multi-sequence fusion further improved predictive performance [31–33], with the combination of T1WI, CE-T1WI, and T2-FLAIR achieving optimal results. This improvement likely arose from the complementary information provided by each sequence [34]. T1WI offers excellent structural detail. T2-FLAIR clearly delineates areas of peritumoral edema. CE-T1WI highlights regions of active tumor enhancement, indicating disruption of the blood-brain barrier and reflecting changes in the tumor microenvironment.

SHAP analysis identified the top-5 radiomic features contributing to the DMSFN model’s performance across the three tasks. The most important feature, “Skewness”, measures asymmetry of the value distribution relative to the mean, reflecting deviations from normality [35]. Another key feature, GLSZM-derived “ZoneEntropy”, quantifies the randomness of region sizes and grayscale levels, with higher values reflecting greater textural heterogeneity [36]. Higher values of both features were observed in HGG. Similarly, IDH-WT glioma exhibited higher “Skewness” values, reflecting the invasive tumor behavior and heterogeneous microenvironments [37, 38].

This study has several limitations. First, inherent class imbalance exists due to the nature of the disease and was mitigated during training using SMOTE. Second, distribution heterogeneity across multi-center cohorts may have contributed to performance degradation. Although conventional preprocessing was applied, residual domain shifts may have affected model generalizability. Future work should explore larger multi-center cohorts and advanced domain generalization or adaptation techniques to enhance model robustness. Finally, tumor habitats in this study were derived solely from ADC maps, reflecting tissue cellularity. Given that tumor heterogeneity encompasses multiple biological dimensions, integrating additional physiological MRI modalities, such as DSC-MRI, arterial spin labeling, or blood oxygen level-dependent imaging, may enable more comprehensive and biologically informative multiparametric habitat models.

Conclusions

This study introduced an MRI radiomics approach utilizing ADC-derived cellularity habitats for ADG grading and IDH mutation status prediction. By focusing on tumor habitat regions, the method significantly improved both model interpretability and predictive accuracy compared to traditional whole-tumor analysis. The results underscore the value of radiomic features derived from tumor habitats in capturing intratumoral heterogeneity and characterizing the tumor microenvironment. Furthermore, the integration of multiparametric radiomics data holds promising potential for advancing clinical applications and enhancing diagnostic precision. This work not only expands the scope of radiomics in assessing tumor heterogeneity but also offers a practical perspective for developing personalized treatment strategies.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

MRI

Magnetic resonance imaging

CNS

Central nervous system

WHO

World Health Organization Classification

ADG

Adult-type diffuse glioma

LGG

Low-grade glioma

HGG

High-grade glioma

IDH

Isocitrate dehydrogenase

IDH-WT

IDH-wildtype

IDH-MT

IDH-mutant

UCSF-PDGM

The University of California San Francisco Preoperative Diffuse Glioma MRI

TCIA

The Cancer Imaging Archive

VOI

Volume of interest

DWI

Diffusion-weighted imaging

ADC

Apparent diffusion coefficient

DKI

Diffusion kurtosis imaging

DSC

Dynamic susceptibility contrast-enhanced

DMSFN

Disentangled-learning-based multi-sequence fusion network

SSE

Sum of squared errors

GLCM

Gray-level co-occurrence matrix

GLRLM

Gray-level run-length matrix

GLSZM

Gray-level size-zone matrix

GLDM

Gray-level dependence matrix

NGTDM

Neighboring gray-tone-difference matrix

SMOTE

Synthetic Minority Over-sampling Technique

ROC

Receiver operating characteristic

AUC

Area under the curve

ACC

Accuracy

SEN

Sensitivity

SPE

Specificity

FDR

False discovery rate

SHAP

Shapley Additive exPlanation

Author contributions

Fangrong Liang and Ruimeng Yang had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Material preparation and data collection and analysis were performed by Jiaxin Lin, Junjie Li, Jie Zhan, Ruili Wei, Wanli Zhang, Yujie He, and Mengni Wu. The first draft of the manuscript was written by Fangrong Liang, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This work was funded by the National Natural Science Foundation of China (82371908, 81971574), the Guangdong Natural Science Foundation (2024A1515012177), the Guangdong Basic and Applied Basic Research Foundation (2021A1515220060), Guangzhou Key Laboratory of Molecular Imaging and Clinical Translational Medicine (202201020376), Medical Science and Technology Research Project of Guangdong Province (A2024637).

Data availability

Institutional data is unavailable due to privacy or ethical restrictions but can be requested from the corresponding author. Public data (UCSF-PDGM) is available in TCIA (DOI: 10.7937/ tcia.bdgf-8v37).

Declarations

Ethics approval and consent to participate

Due to the retrospective nature of the study and the use of de-identified data, patient informed consent was waived. Approval was granted by the Ethics Committee of Beijing Tiantan Hospital (Approval No. KY2022-078-04). The study was conducted in accordance with the ethical principles of the Declaration of Helsinki.

Consent for publication

De-identified data were collected with a waiver of written informed consent.

Competing interests

The authors declare no competing interests.