Application of a multimodal MRI model integrating radiomics and habitat features for predicting glioma pathology and prognosis

Abstract

Background

Accurate grading and prognostic assessment of glioma requires integrating key molecular biomarkers, including IDH mutation status and the Ki-67 proliferation index. However, current radiomics studies often focus on single-task predictions and rely on manual tumor segmentation, which fails to capture intratumoral spatial heterogeneity. This study proposes an automated whole-tumor segmentation-based multimodal MRI approach integrating habitat radiomics to achieve noninvasive, multitask prediction of WHO grade, IDH mutation, Ki-67 labeling index (LI), and 2-year postoperative survival in glioma.

Methods

This retrospective study enrolled 185 patients with pathologically confirmed glioma. Preoperative multimodal MRI - including T1-weighted imaging (T1WI), T2-weighted fluid-attenuated inversion recovery (T2W-FLAIR), and T1-weighted contrast-enhanced imaging (T1W CE) - was acquired for analysis. Using the uAI Research Portal platform, we performed automated whole-tumor segmentation and subsequent feature extraction, deriving 2,264 radiomics features and 61 habitat-based features. Predictive models were developed using multiple machine learning algorithms, and feature selection was rigorously performed within the training folds of a five-fold cross-validation to prevent overfitting. Model performance was evaluated using AUC, accuracy, sensitivity, and specificity, with statistical comparisons conducted performed DeLong’s test.

Results

The habitat model exhibited superior sensitivity in capturing tumor heterogeneity across all four prediction tasks. Building on this, the integrated model combining habitat and conventional radiomics features, achieved the highest overall predictive performance, with AUCs of 0.916 (95% CIs: 0.858–0.975) for glioma grading, 0.877 (95% CIs: 0.828–0.926) for IDH mutation status, 0.859 (95% CIs: 0.788–0.930) for Ki-67 LI, and 0.906 (95% CIs: 0.837–0.974) for 2-year survival prediction, consistently outperforming single-modality models. SHAP interpretability analysis revealed that patient age exhibited strong correlation with tumor grade, IDH mutation status, and Ki-67 LI. Furthermore, tumor grade, IDH status, and Ki-67 LI demonstrated potential predictive value for 2-year postoperative survival.

Conclusions

The automated habitat radiomics framework effectively quantified intratumoral spatial heterogeneity in glioma. When combined with conventional radiomics, it significantly enhanced accuracy in predicting key molecular and clinical endpoints.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-026-02396-y.

Introduction

The 2021 WHO Classification of Central Nervous System Tumors (5th edition, WHO CNS5) mandates comprehensive glioma grading across tiers 1–4 through integrated assessment of histological features and molecular biomarkers, marking a fundamental paradigm shift toward integrated molecular-pathological diagnostics [1]. This transformation underscores the indispensable role of key molecular markers, including isocitrate dehydrogenase (IDH) gene mutation status, 1p/19q codeletion, and Ki-67 labeling index (LI) - in defining tumor classification and grading. For instance, IDH mutation status not only critically differentiates diffuse astrocytomas from glioblastomas but also significantly correlates with therapeutic sensitivity and long-term survival [2, 3]. Concurrently, Ki-67 LI serves as a reliable indicator of cellular proliferative activity, with elevated expression demonstrating positive correlation with tumor malignancy and providing crucial parameters for assessing aggressiveness [4, 5]. Furthermore, 1p/19q co-deletion in oligodendrogliomas has emerged as a pivotal genetic determinant guiding therapeutic selection, with its presence associated with enhanced chemosensitivity and significantly prolonged overall/progression-free survival [6]. Despite this diagnostic revolution, clinical management of glioma remains profoundly challenging. Its highly invasive nature and marked intratumoral heterogeneity contribute to universally poor prognoses, particularly for high-grade gliomas (HGGs) where median survival persists at approximately 15 months even with aggressive treatment. This stark reality underscores the urgent need for more precise, noninvasive diagnostic and prognostic tools.

Radiomics technology addresses this need through high-throughput analysis of conventional medical imaging (e.g., MRI), systematically extracting quantitative features that encompass tumor morphology, size, volume, intensity, and texture characteristics as noninvasive biomarkers reflecting underlying biological information [7]. This approach has demonstrated considerable potential in early glioma detection, molecular subtyping, grading assessment, and treatment monitoring [8, 9]. However, conventional methods that analyze tumors as a whole fail to capture their defining characteristic: spatial intratumoral heterogeneity. For example, T1W CE regions reflect blood-brain barrier disruption, where maximal resection significantly prolongs overall survival (OS) [10], while glioblastomas with Ki-67 LI exceeding 20% typically exhibit higher microvascular density and consequently more pronounced enhancement [11]. To better characterize this heterogeneity, tumor habitat analysis partitions tumors and peritumoral regions into functionally and pathologically distinct subregions [12]. By analyzing pixel intensity and texture variations, this method enhances understanding spatial distribution of the tumor microenvironment and effectively predicts key pathological features [13]. Zhu et al. [14] demonstrated this in a multicenter study of 387 primary glioma patients, combining deep learning with habitat radiomics to develop models that accurately predicting tumor grade (validation AUC = 0.938, test AUC = 0.879), Ki-67 LI (validation AUC = 0.950, test AUC = 0.813), and IDH status (validation AUC = 0.937, test AUC = 0.816). Complementary studies that employed diffusion kurtosis imaging (DKI), dynamic susceptibility contrast MRI (DSC-MRI), and dynamic contrast-enhanced MRI (DCE-MRI) parameters have achieved interpretable results in predicting glioma characteristics [15–17], and confirmed that integrated approaches yield superior predictive performance. Nevertheless, existing researches remain fragmented, with few simultaneously capturing global tumor characteristics and intratumoral heterogeneity or integrating multimodal information to construct predictive models. Moreover, most investigations typically focus on single imaging modalities or isolated clinical endpoints (e.g., grading or survival prediction alone), thereby constraining model comprehensiveness and failing to fully leverage multimodal data or adequately address tumor heterogeneity.

To overcome these limitations, our study employed multiparametric MRI sequences (T1WI, T2W-FLAIR, T1W CE) using unified whole-tumor automated segmentation and habitat radiomics analysis to construct interpretable machine learning models. This framework enabled simultaneous multitask assessment of four core clinical parameters: WHO grade, IDH status, Ki-67 LI, and 2-year postoperative survival status.

Materials and methods

Patients

All patients diagnosed with glioma at Huadong Hospital, Fudan University, from 2017 to 2024 were screened, and the enrollment flowchart is presented in Fig. 1. Inclusion criteria were: (1) Histologically confirmed glioma cases; (2) completion of imaging examinations within 4 weeks prior to surgery; and (3) availability of complete pathological data. Exclusion criteria included: (1) Incomplete MRI data; (2) poor-quality MRI images; and (3) concurrent organic central nervous system disorders (e.g., cerebrovascular events). Follow-up was conducted by reviewing patient return visit records in the electronic medical record system and PACS, supplemented by telephone interviews. The follow-up cutoff date was November 27, 2024, with a minimum follow-up duration of 24 months. This study complied with the Declaration of Helsinki and was approved by the Ethics Review Committee of Huadong Hospital, Fudan University (Ethics approval number: 2023K106).

Fig. 1.

Flowchart of enrollment

According to the WHO CNS5 criteria [1], tumors were classified into grades 1–4, with grades 1–2 designated as low-grade gliomas (LGGs) and grades 3–4 as HGGs. Immunohistochemical analysis was performed to assess IDH mutation status and Ki-67 LI in tumor specimens. Based on our team’s previous study [18], Ki-67 LI was dichotomized into high and low levels using a cutoff threshold of 9.5%.

MRI acquisition

MRI examinations were performed using a Siemens 3.0 T superconducting MR scanner (MAGNETOM Vida/Skyra/Prisma/Verio, Siemens Healthineers, Erlangen, Germany) equipped with a 20-channel head-neck combined coil. The imaging protocol included axial T1WI, T2W-FLAIR, and T1W CE, with detailed scanning parameters provided in the Supplementary Table 1.

Image processing and habitat construction

MRI image processing, habitat analysis, and model construction were performed using the uAI Research Portal (United Imaging Intelligence, China) [19–22].

Tumor segmentation

Initial preprocessing involved N4 bias field correction of the original MRI data. Subsequently, rigid registration was performed using ANTs to align T1W CE and T2W-FLAIR images with T1WI. Following skull stripping, adaptive grayscale normalization was applied to enhance data consistency. The uAI whole-tumor segmentation model (trained on the BraTS 2021 dataset using VB-Net architecture) was employed to delineate volumetric regions of interest (VOIs) [23, 24]. This delineation encompassed all relevant glioma subregions, including enhancing tumor, the peritumoral edematous/invaded tissue and the necrotic tumor core. The description regarding this model’s segmentation capability for non-enhancing gliomas was provided in the Supplementary Note. All segmentation results underwent visual inspection by an experienced clinical radiologist, with manual corrections applied to any erroneous segmentations.

Whole-tumor region radiomics feature extraction

Radiomics features were comprehensively and systematically extracted across T1WI, T2W-FLAIR, T1W CE and their derived images generated through various preprocessing techniques including filtered transformations (multi-scale LoG), wavelet transforms (Wavelet), BoxMean, BinomialBlur, CurvatureFlow, DiscreteGaussian, RecursiveGaussian, LaplacianSharpening, noise addition (additive Gaussian noise, speckle noise, shot noise), and normalization processing. Seven major categories of radiomics features were extracted (first-order, shape, GLCM, GLRLM, GLSZM, GLDM, NGTDM), yielding total of 2264 radiomics features from each sequence.

Habitat analysis and multidimensional feature extraction

Voxel-wise features derived from multiparametric MRI (T1WI/T2W-FLAIR/T1W CE) were subjected to unsupervised k-means clustering for automated habitat subregion partitioning. The elbow method dynamically determined the optimal cluster number (k). Mean signal intensities of each subregion category across three MRI sequences were computed to construct group-level subregional feature distribution atlases. By mapping tumor voxels of individual patients to population-derived cluster centroids, personalized subregional spatial distribution maps were generated, from which quantitative features were extracted in across three dimensions: (1) subregional volumetric features: evaluating dominant habitat subregions within tumors by calculating absolute volumes (voxel counts per subregion) and relative volumes (percentage of subregional volume relative to total tumor volume); (2) signal intensity features: Characterizing typical imaging manifestations and internal signal homogeneity of habitats through mean signal intensities and standard deviations per MRI sequence for each subregion; (3) tumor spatial heterogeneity features: (i) constructing of multiregional spatial interaction (MSI) matrices to quantify spatial arrangement patterns and feature disparities among subregions. Following Wu et al. [25], spatial adjacency relationships were modeled via graph construction, where nodes represented subregions and edges connecting adjacent nodes (boundary-overlapping subregions) were assigned unity values. Twenty-two radiomics features (18 first-order and 4 s-order statistical features) were subsequently extracted from MSI matrices to quantify intratumoral spatial heterogeneity. (ii) quantification of intratumoral inhomogeneity and diversity through ITHscore calculations. Based on Li et al.‘s [26] multiscale radiomics approach, ITHscore (range: 0–1) was computed based on connected component numbers (n) and maximal connected component areas (S i, max) per subregion (k), where elevated ITHscore indicated more fragmented tumor architecture reflecting greater cellular compositional and spatial distribution heterogeneity, yielding seven additional features.

Model construction

For all four clinical tasks, we developed seven distinct models using a rigorously validated five-fold cross-validation approach with embedded Synthetic Minority Over-sampling Technique (SMOTE): (a) single T1WI radiomics model; (b) single T2W-FLAIR radiomics model; (c) single T1W CE radiomics model; (d) trimodal fused radiomics model (radiomics_fusion); (e) habitat feature model (habitat); (f) radiomics-habitat feature fusion model (radiomics_habitat); and (g) comprehensive fusion model incorporating clinical information (clinical_radiomics_habitat). For predictive models of glioma grade, IDH mutation status, and Ki67 levels, clinical information comprised age and sex. For survival prediction, clinical variables encompassed age, sex, resection extent, postoperative treatment modalities, tumor grade, IDH mutation status, Ki67 levels, and 1p/19q status.

The pipeline was carefully designed to prevent data leakage: for each fold, SMOTE was applied exclusively to the training set (four folds) prior to feature selection, with the original class distribution being preserved in the test fold. Feature selection was performed independently on each SMOTE-augmented training subset, and the most frequently selected features across all folds were aggregated to form a consensus feature set. Models were subsequently trained on the SMOTE-balanced training data using this consensus feature set and evaluated on the original, non-augmented test fold. This process was repeated across all folds, with final performance metrics representing averages from all test folds. Predictive models were constructed using 13 machine learning algorithms provided by the uAI Research Portal platform, including adaptive boosting, bagging decision tree, decision tree, Gaussian process, gradient boosting decision tree, K-nearest neighbors, random forest, logistic regression, extreme gradient boosting, stochastic gradient descent, support vector machine, quadratic discriminant analysis, partial least squares-discriminant analysis. To ensure a fair comparison and avoid overfitting, we performed hyperparameter tuning for each algorithm using a grid search strategy within the inner loop of the five-fold cross-validation. The performance of each set of hyperparameters was evaluated based on the AUC in the validation fold. Ultimately, for each prediction task, the algorithm (with its optimal hyperparameters) yielding the highest mean validation AUC across the five folds was selected as the final predictive model. Additionally, fusion models were specifically built at the feature level.

Model diagnostic performance was evaluated using AUC with 95% confidence intervals (CIs), sensitivity, specificity, accuracy, precision, and F1-score. Receiver operating characteristic (ROC) curves were used to assess discriminative performance, calibration curves were used to visualize the agreement between predicted and actual probabilities, and decision curve analysis (DCA) quantified clinical net benefits across different decision thresholds. SHAP analysis was applied to quantify the contribution of each feature to model predictions.

Statistical analysis

Statistical analyses were performed using SPSS 26.0.0. The Shapiro-Wilk test was used to assess normality of continuous variables. Normally distributed continuous variables were expressed as mean ± standard deviation and compared using independent-samples t-tests; non-normally distributed continuous variables were expressed as median (interquartile range) and compared using the nonparametric Mann-Whitney U test. Categorical variables were described using numbers and percentages (%) and compared via chi-square test. The DeLong test was employed to compare predictive performance between different models. A P value < 0.05 was considered statistically significant.

Results

Study cohorts

A total of 185 patients were included in the analysis, comprising 141 HGGs and 44 LGGs. The baseline characteristics of the two patient groups are presented in Table 1. No significant difference was observed in gender distribution between the HGGs and LGGs groups (P = 0.458). The age of HGGs group (57.2 ± 12.7 years) was significantly higher than that of the LGGs group (39.9 ± 13.3 years) (P < 0.001). Regarding IDH wild-type distribution, the HGGs group showed a higher proportion (HGGs: 80.9% vs. LGGs: 45.5%, P < 0.001). Additionally, Ki-67 LI in the HGGs group [25.0% (15.0%, 40.0%)] was significantly higher than in LGGs group [3.0% (1.3%, 5.0%)] (P < 0.001).

Table 1.

Baseline characteristics of enrolled patients

		All (n = 185)	HGGs (n = 141)	LGGs (n = 44)	χ2/t/Z	P
Gender	Male	118(63.8%)	92(65.2%)	26(59.1%)	0.550	0.458
	Female	67(36.2%)	49(34.8%)	18(40.9%)
Age at diagnosis (y)		53.1 ± 14.8	57.2 ± 12.7	39.9 ± 13.3	-7.794	< 0.001
Molecular subtype	IDHwt	134(72.4%)	114(80.9%)	20(45.5%)	21.042	< 0.001
	IDHmut	51(27.6%)	27(19.1%)	24(54.5%)
Ki-67 LI		20.0%(8.0%, 30.0%)	25.0% (15.0%, 40.0%)	3.0%(1.3%, 5.0%)	-9.368	< 0.001

Note: Statistical significance was defined as P < 0.05, with significant results indicated in bold type

Habitat parameter quantification

The optimization process for determining cluster number (k) using the elbow method was illustrated in Fig. 2A. A distinct inflection point (elbow) was observed at k = 4, which was consequently selected as the optimal cluster number that balanced model complexity and explanatory power. Ultimately, each case yielded 61 habitat radiomics features, calculated as k + k + sequence number (3) × k × 2 + 22 + 7 = 61. Figure 2B demonstrates the three-dimensional clustering results of feature dimensions. Figure 2C presents exemplary images of habitat subregion partitioning based on the elbow method.

Fig. 2.

Cluster number selection and habitat subregion partitioning. (A) Elbow method plot. (B) Three-dimensional visualization of habitat feature clustering. (C) Exemplary images of habitat partitioning. Four distinct subregions were delineated based on k = 4. T1WI, T1-weighted imaging; T2W-FLAIR, T2-weighted fluid-attenuated inversion recovery; T1W CE, T1-weighted contrast-enhanced imaging

Results of glioma grade classification

Table 2 presents the classification performance of different models for glioma grade. Figure 3 displays performance evaluation plots for each predictive model in the training set (Fig. 3A ~ C) and testing set (Fig. 3D ~ F). SHAP plots illustrating feature importance for the optimal models are presented in Fig. 3G-H. Comparative analyses of model performance are provided in Supplementary Fig. 1. The radiomics model demonstrated superior predictive performance compared to the habitat model. Comprehensive evaluation across all metrics indicated that the clinical_radiomics_habitat model achieved optimal predictive performance with favorable accuracy and clinical utility. In the training set, the model yielded an accuracy of 0.887, sensitivity of 0.954, specificity of 0.669, precision of 0.904, F1 score of 0.928, and AUC of 0.941 (95% CIs: 0.915–0.966). In the test set, performance metrics were as follows: accuracy 0.870, sensitivity 0.943, specificity 0.628, precision 0.893, F1 score 0.917, and AUC 0.916 (95% CIs: 0.858–0.975). Furthermore, DeLong tests(Supplementary Fig. 1) demonstrated that the clinical_radiomics_habitat model achieved significantly higher AUC values than any single-modality radiomics model as well as the habitat model (P < 0.05).

Table 2.

Predictive performance of different models for glioma grade

	Model	Feature number	Accuracy	Sensitivity	Specificity	Precision	F1-score	AUC	AUC 95%CI
Training	T1WI	8	0.870	0.989	0.477	0.861	0.921	0.866	0.830–0.902
	T2W-FLAIR	7	0.860	0.968	0.506	0.865	0.914	0.904	0.873–0.935
	T1W CE	13	0.918	0.924	0.901	0.968	0.946	0.974	0.957–0.991
	radiomics_fusion	5	0.863	0.941	0.605	0.886	0.913	0.894	0.861–0.926
	habitat	6	0.844	0.952	0.488	0.859	0.903	0.888	0.855–0.921
	radiomics_habitat	6	0.878	0.940	0.674	0.904	0.922	0.913	0.883–0.943
	clinical_radiomics_habitat	6	0.887	0.954	0.669	0.904	0.928	0.941	0.915–0.966
Testing	T1WI	8	0.788	0.936	0.302	0.815	0.871	0.798	0.714–0.883
	T2W-FLAIR	7	0.799	0.936	0.349	0.825	0.877	0.817	0.736–0.899
	T1W CE	13	0.832	0.879	0.674	0.899	0.889	0.851	0.776–0.927
	radiomics_fusion	5	0.864	0.943	0.605	0.887	0.914	0.878	0.809–0.947
	habitat	6	0.810	0.936	0.395	0.835	0.883	0.834	0.755–0.913
	radiomics_habitat	6	0.859	0.936	0.605	0.886	0.910	0.888	0.821–0.955
	*clinical_radiomics_habitat	6	0.870	0.943	0.628	0.893	0.917	0.916	0.858–0.975

*Among the clinical features, only age was ultimately retained, as gender was not selected by the feature selection algorithm for inclusion in the final model

Fig. 3.

Evaluation of predictive performance for glioma grade by different models. (A) ROC curve of the training set. (B) Calibration curve of the training set. (C) DCA of the training set. (D) ROC curve of the testing set. (E) Calibration curve of the testing set. (F) DCA of the testing set. (G) Mean absolute SHAP value plot for the optimal model. (H) Five-fold cross-validation SHAP plot for the optimal model

Classification results of IDH mutation status in glioma

Table 3 shows the classification performance of different models for IDH mutation status in glioma. Figure 4 displays performance evaluation plots for each predictive model in the training set (Fig. 4A ~ C) and testing set (Fig. 4D ~ F). SHAP plots illustrating feature importance for the optimal models are presented in Fig. 4G-H. The radiomics model demonstrated superior predictive performance compared to the habitat model. Comparative analyses of model performance are provided in Supplementary Fig. 2. Comprehensive evaluation across all metrics indicated that the clinical_radiomics_habitat model achieved optimal predictive performance with favorable accuracy and clinical utility. In the training set, the model yielded an accuracy of 0.802, sensitivity of 0.485, specificity of 0.923, precision of 0.707, F1 score of 0.576, and AUC of 0.819 (95% CIs: 0.789–0.849). In the test set, performance metrics were as follows: accuracy 0.793, sensitivity 0.412, specificity 940, precision 0.724, F1 score 0.525, and AUC 0.877 (95% CIs: 0.828–0.926). Furthermore, DeLong tests(Supplementary Fig. 2) demonstrated that the clinical_radiomics_habitat model achieved significantly higher AUC values than any single-modality radiomics model as well as the habitat model (P < 0.05).

Table 3.

Predictive performance of different models for IDH mutation status in glioma

	Model	Feature number	Accuracy	Sensitivity	Specificity	Precision	F1-score	AUC	AUC 95%CI
Training	T1WI	6	0.815	0.515	0.930	0.739	0.607	0.818	0.787–0.848
	T2W-FLAIR	7	0.791	0.451	0.921	0.687	0.544	0.800	0.768–0.832
	T1W CE	13	0.889	0.667	0.974	0.907	0.768	0.944	0.929–0.960
	radiomics_fusion	7	0.838	0.740	0.876	0.696	0.717	0.883	0.859–0.906
	habitat	5	0.754	0.289	0.932	0.621	0.395	0.778	0.745–0.812
	radiomics_habitat	8	0.856	0.676	0.925	0.775	0.723	0.870	0.845–0.895
	clinical_radiomics_habitat	8	0.802	0.485	0.923	0.707	0.576	0.819	0.789–0.849
Testing	T1WI	6	0.793	0.471	0.917	0.686	0.558	0.788	0.722–0.854
	T2W-FLAIR	7	0.750	0.392	0.887	0.571	0.465	0.745	0.672–0.818
	T1W CE	13	0.793	0.490	0.910	0.676	0.568	0.807	0.744–0.870
	radiomics_fusion	7	0.799	0.667	0.850	0.630	0.648	0.831	0.773–0.890
	habitat	5	0.755	0.314	0.925	0.615	0.416	0.772	0.703–0.840
	radiomics_habitat	8	0.859	0.667	0.932	0.791	0.723	0.852	0.798–0.906
	clinical_radiomics_habitat	8	0.793	0.412	0.940	0.724	0.525	0.877	0.828–0.926

Fig. 4.

Evaluation of predictive performance for IDH mutation status in glioma by different models. (A ~ C) Training set. (D ~ F) Testing set. (G) Mean absolute SHAP value plot for the optimal model. (H) Five-fold cross-validation SHAP plot for the optimal model

Classification results of Ki-67 LI levels in glioma

Table 4 demonstrates the classification performance of different models for Ki-67 LI levels in glioma. Figure 5 presents performance evaluation plots for each predictive model in the training set (Fig. 5A ~ C) and testing set (Fig. 5D ~ F). SHAP plots illustrating feature importance for the optimal models are presented in Fig. 5G-H. Comparative analyses of model performance are provided in Supplementary Fig. 3. The radiomics model exhibited superior predictive performance compared to the habitat model. Comprehensive evaluation across all metrics indicated that the clinical_radiomics_habitat model achieved optimal predictive performance with favorable accuracy and clinical utility. In the training set, the model yielded an accuracy of 0.836, sensitivity of 0.912, specificity of 0.612, precision of 0.873, F1 score of 0.892, and AUC of 0.870 (95% CIs: 0.835–0.904). In the test set, performance metrics were as follows: accuracy 0.821, sensitivity 0.898, specificity 0.596, precision 0.866, F1 score 0.882, and AUC 0.859 (95% CIs: 0.788–0.930). Furthermore, DeLong tests(Supplementary Fig. 3)demonstrated that the clinical_radiomics_habitat model achieved significantly higher AUC values than any single-modality radiomics model as well as the radiomics_fusion model (P < 0.05).

Table 4.

Predictive performance of different models for Ki-67 LI in glioma

	Model	Feature number	Accuracy	Sensitivity	Specificity	Precision	F1-score	AUC	AUC 95%CI
Training	T1WI	6	0.840	0.905	0.649	0.883	0.894	0.890	0.858–0.922
	T2W-FLAIR	4	0.799	0.993	0.234	0.791	0.880	0.677	0.630–0.724
	T1W CE	9	0.859	0.954	0.580	0.869	0.910	0.872	0.838–0.906
	radiomics_fusion	4	0.844	0.964	0.495	0.848	0.902	0.855	0.819–0.891
	habitat	6	0.789	0.894	0.484	0.835	0.863	0.814	0.774–0.853
	radiomics_habitat	9	0.708	0.668	0.824	0.917	0.773	0.844	0.807–0.881
	clinical_radiomics_habitat	6	0.836	0.912	0.612	0.873	0.892	0.870	0.835–0.904
Testing	T1WI	6	0.766	0.883	0.426	0.818	0.849	0.772	0.686–0.857
	T2W-FLAIR	4	0.788	0.993	0.191	0.782	0.875	0.616	0.519–0.712
	T1W CE	9	0.815	0.927	0.489	0.841	0.882	0.796	0.714–0.877
	radiomics_fusion	4	0.826	0.942	0.489	0.843	0.890	0.816	0.738–0.895
	habitat	6	0.799	0.905	0.489	0.838	0.870	0.806	0.726–0.886
	radiomics_habitat	9	0.712	0.679	0.809	0.912	0.778	0.844	0.770–0.918
	clinical_radiomics_habitat	6	0.821	0.898	0.596	0.866	0.882	0.859	0.788–0.930

Fig. 5.

Evaluation of predictive performance for Ki-67 LI in glioma by different models. (A ~ C) Training set. (D ~ F) Testing set. (G) Mean absolute SHAP value plot for the optimal model. (H) Five-fold cross-validation SHAP plot for the optimal model

Prediction of 2-year survival status after glioma surgery

Supplementary Table 2 presents baseline characteristics of patients enrolled in survival analysis. Patients in the 2-year postoperative mortality group exhibited significantly higher age, tumor grade, and Ki-67 LI levels compared to the survival group (all P < 0.001). The mortality group demonstrated higher proportions of IDH wild-type (57 (87.7%) vs. 16 (43.2%), P < 0.001) and 1p/19q non-codeletion (41 (93.2%) vs. 14 (51.9%), P < 0.001) relative to the survival group. Table 5 displays predictive performance of different models for 2-year survival status in glioma. Figure 6 illustrates performance evaluation plots for each prediction model in training (Fig. 6A ~ C) and testing sets (Fig. 6D ~ F). SHAP plots illustrating feature importance for the optimal models are presented in Fig. 6G-H. Comparative analyses of model performance are provided in Supplementary Fig. 4. The radiomics model outperformed the habitat model in predictive efficacy. Comprehensive evaluation across all metrics indicated that the clinical_radiomics_habitat model achieved optimal predictive performance with favorable accuracy and clinical utility: training set - accuracy = 0.914; sensitivity = 0.969; specificity = 0.818; precision = 0.903; F1-score = 0.935; AUC = 0.950 (95% CIs: 0.925 ~ 0.975); testing set - accuracy = 0.863; sensitivity = 0.908; specificity = 0.784; precision = 0.881; F1-score = 0.894; AUC = 0.906 (95% CIs: 0.837 ~ 0.974). Furthermore, DeLong tests(Supplementary Fig. 4)demonstrated that the clinical_radiomics_habitat model achieved significantly higher AUC values than the T1W CE, T2W-FLAIR, and habitat models (P < 0.05).

Table 5.

The predictive performance of different models for 2-year survival status in glioma

	Model	Feature number	Accuracy	Sensitivity	Specificity	Precision	F1-score	AUC	AUC 95%CI
Training	T1WI	8	0.772	0.865	0.608	0.795	0.829	0.851	0.809–0.893
	T2W-FLAIR	8	0.811	0.942	0.581	0.798	0.864	0.893	0.857–0.929
	T1W CE	9	0.801	0.842	0.730	0.846	0.844	0.868	0.828–0.908
	radiomics_fusion	12	0.865	0.954	0.709	0.852	0.900	0.943	0.916–0.970
	habitat	6	0.904	1.000	0.736	0.870	0.930	0.995	0.986–1.003
	radiomics_habitat	7	0.762	0.781	0.730	0.835	0.807	0.853	0.811–0.894
	clinical	5	0.877	0.919	0.804	0.892	0.905	0.915	0.882–0.947
	clinical_radiomics_habitat	7	0.914	0.969	0.818	0.903	0.935	0.950	0.925–0.975
Testing	T1WI	8	0.755	0.877	0.541	0.770	0.820	0.808	0.715–0.902
	T2W-FLAIR	8	0.755	0.892	0.514	0.763	0.823	0.781	0.682–0.879
	T1W CE	9	0.725	0.754	0.676	0.803	0.778	0.837	0.749–0.925
	radiomics_fusion	12	0.775	0.877	0.595	0.792	0.832	0.824	0.733–0.914
	habitat	6	0.647	0.954	0.108	0.653	0.775	0.652	0.538–0.765
	radiomics_habitat	7	0.745	0.769	0.703	0.820	0.794	0.831	0.742–0.920
	clinical	5	0.824	0.877	0.730	0.851	0.864	0.862	0.781–0.944
	clinical_radiomics_habitat	7	0.863	0.908	0.784	0.881	0.894	0.906	0.837–0.974

Fig. 6.

Evaluation of predictive performance for 2-year survival status in glioma by different models. (A ~ C) Training set. (D ~ F) Testing set. (G) Mean absolute SHAP value plot for the optimal model. (H) Five-fold cross-validation SHAP plot for the optimal model

Disscusion

This study applies habitat-based radiomics approaches to address four key prediction tasks in glioma simultaneously: WHO grade, IDH mutation status, Ki-67 LI levels, and postoperative survival. Unlike traditional radiomics research, our approach not only quantifies the tumor as a whole but, more importantly, deciphers its internal spatial heterogeneity, providing a powerful, non-invasive lens through which to interpret the biological behavior of glioma, enhancing the explanatory power of imaging.

Glioma, as the most common primary malignant tumor in the intracranial cavity, is characterized by significant spatial and temporal heterogeneity [1]. Different regions within the tumor may exhibit varying cellular density, vascularization degree, necrosis extent, and proliferative activity, with this heterogeneity directly influencing the tumor’s biological behavior, treatment response, and prognosis [27]. Our habitat analysis method employs unsupervised clustering to identify subregions with similar imaging features, with each subregion representing a distinct “microenvironment,” which thereby accurately captures the tumor’s complex biological characteristics. As shown in Fig. 2B, the four habitat subregions exhibit distinct imaging patterns that potentially reflect differential tissue characteristics based on established MRI-pathology correlations: Cluster 1 is characterized by high T2W-FLAIR signal and low T1W CE enhancement, a pattern frequently observed in peritumoral inflammatory edema and infiltration zones; Cluster 2 exhibits low T2W-FLAIR signal, which may correspond to necrotic components commonly seen in gliomas. The combination of hyperintense T1WI/T2W-FLAIR signals with hypointense enhancement in Cluster 3 could be consistent with hemorrhagic or protein-rich fluid regions, whereas Cluster 4, characterized by hyperintensity across all sequences with high T1W CE enhancement, parallels typical imaging features of solid tumor tissue. Nevertheless, we emphasize that these associations are inferred from imaging phenotypes alone and remain hypothetical pending histopathological validation. Indeed, unimodal MRI lacks the specificity to reliably discriminate among these pathological constituents, as distinct pathological processes may yield convergent imaging appearances. Consequently, the precise biological significance of each subregion requires verification through emerging techniques such as spatial transcriptomics or whole-slide digital pathology in correlative studies.

At the methodological level, this study adopted a multi-level, multi-dimensional feature extraction and integration strategy with significant advantages. First, extracted global radiomics features of the tumor included shape, texture, and signal intensity characteristics, reflecting the macroscopic morphological and signal properties of the tumor. Second, subregional features obtained through habitat analysis could quantify the heterogeneous distribution within the tumor, including the volumetric proportions, spatial distribution patterns, and interrelationships of different subregions. Finally, incorporation of age and sex further enhanced the clinical applicability and predictive accuracy of the model. Regarding feature selection and model construction, robust statistical methods were employed to handle the high-dimensional feature space. Radiomics research often faces the “curse of dimensionality” problem, where the number of features far exceeds the number of samples, making the model highly prone to overfitting. Through appropriate feature selection strategies and cross-validation approaches, we improved the model’s generalization capability. For the prediction of 2-year survival status, while our clinical_radiomics_habitat model demonstrated robust stability (testing AUC = 0.906, Table 5) compared to the severely overfit habitat model (training AUC = 0.995 vs. testing AUC = 0.652, Table 5), concerns regarding potential over-reliance on clinical variables merit clarification. SHAP analysis (Fig. 6G) of the integrated prognostic model identified the following top contributors: (a) Grade, (b) IDH status, (c) Wavelet_firstorder_wavelet-LHH-Mean_t1c, (d) 1p19q codeletion, (e) Log_glrlm_log-sigma-4-0-mm-3D-ShortRunHighGrayLevelEmphasis_flair, (f) Log_glszm_log-sigma-4-0-mm-3D-SizeZoneNonUniformityNormalized_t1c, and (g) Subregion2_z_std. This ranking underscores that imaging features capture distinct tumor biological characteristics beyond conventional clinical variables. Although the DeLong test indicated that its AUC value showed no statistically significant improvement over the clinical model (0.862 vs. 0.906, P > 0.05, Supplementary Fig. 4B), the absolute gain of 0.044 aligned with the following clinically relevant observations: (a) A clinically meaningful increase in specificity (0.730 vs. 0.784, Table 5), thereby enhancing negative predictive ability; and (b) on the DCA plot (Fig. 6F), the clinical_radiomics_habitat model curve consistently occupied the uppermost position, indicating superior clinical net benefit across various threshold probabilities.

The results of this study demonstrated that the optimal feature combinations vary across different prediction tasks, effectively reflecting the complex correlations between various biological indicators and imaging phenotypes. WHO grade prediction might rely more heavily on texture features that reflect cellular density and vascularization degree, as HGGs typically exhibit higher cellular density and richer vascular networks. The SHAP analysis plot similarly revealed that, apart from age, MSI features representing intratumoral texture heterogeneity exerted significant influence on model decision-making. IDH mutation might be associated with the specific metabolic profile, inhibiting tumor angiogenesis and promoting vascular normalization. Thus, the vascular architecture might appear relatively regular, manifesting as weaker and more homogeneous enhancement on T1W CE imaging [28–30]. Correspondingly, SHAP analysis results indicated that T1W CE image features carry relatively greater weight in the IDH prediction task. Ki-67 LI, as a proliferation activity indicator, might have imaging correlates including tumor border irregularity, signal heterogeneity, and spatial relationships with necrotic areas. Highly proliferative regions typically correspond to specific subregions in habitat analysis, potentially manifesting as higher T2WI signal intensity, heterogeneous contrast agent uptake, and distinct boundaries with surrounding normal brain tissue. Survival prediction required comprehensive consideration of multiple factors, including overall tumor characteristics, heterogeneity degree, and patient demographics. The SHAP analysis results for each task consistently aligned with these biological characteristics.

Technically, we employed an automated whole-tumor segmentation technique via the uAI platform, which effectively reduced manual workload and inter-observer variability, while SHAP analysis enhanced model interpretability, thereby facilitating clinical translation. The four predictive tasks of this study are expected to provide a complementary non-invasive assessment method for individualized treatment strategies under the WHO CNS 5 framework. During the preoperative evaluation, this model can serve as a non-invasive predictive tool, enabling clinicians to obtain crucial information regarding tumor grade, molecular characteristics, and prognosis prior to surgery. This held significant value for formulating personalized treatment plans, particularly for patients with deep-seated lesions or high surgical risks, where accurate preoperative predictions might help avoid unnecessary surgical hazards. For treatment monitoring, habitat analysis could be employed to evaluate therapeutic responses. From a precision medicine perspective, this study supports image phenotype-based patient stratification strategies. Distinct habitat patterns might correlate with differential treatment sensitivities, establishing a foundation for developing imaging feature-based therapeutic selection algorithms. Future research could explore integrating radiomic features with multi-omics data including genomics and proteomics to construct more comprehensive patient stratification systems.

In recent years, radiomics research on glioma had developed rapidly, yet most studies have focused on single prediction tasks, such as predicting only IDH status or only survival outcomes [31–33]. The multi-task prediction framework employed in this study better aligns with clinical practice needs, as clinicians typically require simultaneous access to multi-dimensional information for treatment planning. Compared to previous studies, our approach demonstrates advantages in several aspects: First, habitat analysis provides more comprehensive information than traditional whole-ROI analysis. Existing research has demonstrated that tumor heterogeneity significantly correlates with prognosis, but effectively quantifying this heterogeneity has remained a methodological challenge [27]. Habitat analysis offers a viable solution by transforming complex spatial heterogeneity into analyzable numerical features [34]. Secondly, we systematically integrated three complementary feature levels: (a) whole-tumor conventional radiomics, (b) habitat-based subregional heterogeneity descriptors, and (c) clinical variables. Across all prediction tasks, the integrated model consistently outperformed single-source counterparts, corroborating findings from previous investigations [35, 36]. Regarding technological development trends, deep learning methods are increasingly applied in medical image analysis. Future research could explore combining traditional machine learning methods with deep learning approaches to capitalize on their respective advantages. While convolutional neural networks can automatically learn imaging features, traditional methods maintain superiority in feature interpretability. The integration of both approaches may yield improved predictive performance and clinical acceptability. Regarding the Ki-67 cutoff selection, the threshold of 9.5% for stratifying Ki-67 expression levels was determined based on ROC curve analysis from our preliminary investigations [18]. Notably, this value approximates the 10% cutoff commonly employed in previous studies [14, 37, 38], which have demonstrated associations between high Ki-67 expression and adverse prognosis. Our derived threshold (9.5%) deviated by merely 0.5% from the established benchmark, well within the acceptable margin of error for pathological interpretation in our specific context. It is important to note that this value is inherently linked to our specific immunohistochemical protocols and patient demographics. Given the inherent inter-laboratory variability in Ki-67 staining and scoring, therefore, this threshold should be treated as a context-dependent reference. Consequently, caution is warranted that external clinical application may necessitate local recalibration or the use of harmonization techniques to ensure diagnostic consistency across different healthcare settings. Multi-center studies are warranted to further validate the generalizability of this threshold across diverse populations.

While this study yielded positive results, several limitations remain that require improvement in future research. First, despite our stringent enrollment criteria ensuring data quality, the single-center sample size of 185 cases might compromise the generalizability of our models. The limited sample size could result in insufficient statistical power and hinder the detection of subtle yet biologically meaningful predictive signals. Furthermore, the single-center design may introduce selection bias (e.g., regional variations in clinical practice). Future investigations should prioritize multi-center validation (involving at least three independent centers with a total sample size of ≥ 500 cases) and incorporate public datasets (such as TCGA database) to enable cross-population validation. Second, the standardization issue in imaging acquisition remains. Variations in scanning parameters across different hospitals and equipment may impact the stability and reproducibility of radiomic features. Although standardized preprocessing methods were employed, systematic inter-device differences remain challenging. Establishing standardized imaging acquisition and processing protocols is crucial for translating radiomics research into clinical applications. Third, concerns remain regarding validation adequacy. While cross-validation methods were used to assess model performance in this study, an independent external validation cohort was lacking. Ideally, models should be validated on completely independent datasets to demonstrate generalization capability. Prospective validation studies are also necessary to evaluate model performance in real-world clinical settings. Fourth, there are temporal considerations. The survival analysis in this study was set at two years, a timeframe selection requiring stronger theoretical justification. Glioma of different grades and molecular subtypes exhibit distinct natural histories, making fixed time windows potentially inflexible. Future research could explore dynamic time windows or survival predictions at multiple time points. Furthermore, we plan to incorporate Cox proportional hazards analysis, Kaplan-Meier survival curves, and risk stratification analyses in our forthcoming multicenter validation study, which would benefit from a larger sample size more amenable to robust time-to-event modeling.

Conclusion

This study successfully applied habitat analysis-based radiomics methods to multi-task prediction in glioma, providing novel technical approaches for precise diagnosis and individualized treatment of glioma. By integrating whole-tumor characteristics, habitat subregional features, and clinical information, our constructed model demonstrated simultaneous predictive capability for WHO grading, IDH typing, Ki67 expression levels, and two-year survival outcomes, and this multi-task prediction framework better aligned with actual clinical needs.

The research results indicated that habitat analysis could effectively capture spatial heterogeneity in glioma, providing more comprehensive information than traditional methods. The optimal feature combinations differed across prediction tasks, reflecting complex correlations between various biological indicators and imaging phenotypes. The incorporation of clinical features further enhanced the model’s predictive performance and practical utility.

Despite limitations including sample size constraints and standardization challenges, this study made significant methodological contributions and provides clinical value for glioma radiomics research. Through future multicenter collaborations, technical improvements, and clinical validation, this approach held promise to become an important auxiliary tool in glioma diagnosis and treatment, advancing the development of precision medicine in brain tumors.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

LI

Labeling index

T1WI

T1-weighted imaging

T2W-FLAIR

T2-weighted fluid-attenuated inversion recovery

T1W CE

T1-weighted contrast-enhanced imaging

WHO CNS

WHO Classification of Central Nervous System Tumors

IDH

Isocitrate dehydrogenase

HGGs

High-grade gliomas

OS

Overall survival

DKI

Diffusion kurtosis imaging

DSC-MRI

Dynamic susceptibility contrast MRI

DCE-MRI

Dynamic contrast-enhanced MRI

LGGs

Low grade gliomas

VOIs

Volumetric regions of interest

SMOTE

Synthetic Minority Over-sampling Technique

CI

Confidence intervals

ROC

Receiver operating characteristic

DCA

Decision curve analysis

Author contributions

SL, GL and XF conceived and presented idea. LS, ND, DY, YL, MY, JL and XZ collected the data. YY, LS and ZC analyzed the data. YY provided statistical guidance. LS, YY and ZC drafted the manuscript. All authors reviewed the manuscript, and SL made corrections to the manuscript. All authors contributed to the article and approved the submitted version. All authors read and approved the final manuscript.

Funding

This study was funded by the Clinical Research and Cultivation Project of Shanghai ShenKang Hospital Development Center (SHDC2022CRT025), the Joint Research Development Project between Shenkang and United Imaging on Clinical Research and Translation (SKLY2022CRT402) and the Huadong Hospital Key Discipline Construction Project (ZDXK2209).

Data availability

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

Declarations

Ethics approval and consent to participate

This retrospective study was approved by the Ethics Committee of Huadong Hospital Affiliated to Fudan University (2023K106), and informed consent was waived, given the retrospective study and anonymised data of patients. This study was conducted in accordance with the World Medical Association Declaration of Helsinki-Ethical Principles for Medical Research Involving Human Subjects.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.