Predicting early recurrence of craniopharyngioma using multi-omics radiomic modeling: a retrospective cohort study

Abstract

Craniopharyngiomas (CPs) are benign tumors arising in the sellar and suprasellar regions, but they frequently recur after surgery because of their proximity to critical neurovascular and hypothalamic structures. This study aimed to develop and validate a habitat-based multimodal radiomics model for early postoperative recurrence prediction in CP. This retrospective single-center cohort included 106 patients with pathologically confirmed CP selected from 136 consecutive surgically treated cases. Preoperative magnetic resonance imaging (MRI) scans (T1-weighted imaging, contrast-enhanced T1-weighted imaging, and T2 -weighted imaging) were used for tumor habitat analysis based on unsupervised clustering of voxel-level radiomic features. Radiomic features were extracted from whole-tumor and habitat-defined subregions, followed by feature selection within the training cohort using univariate filtering, correlation pruning, minimum redundancy maximum relevance (mRMR), and least absolute shrinkage and selection operator (LASSO). Clinical, radiomics, habitat-based, and Combined models. Model performance was evaluated using Receiver Operating Characteristic (ROC), calibration, and decision curve analyses. Shapley Additive exPlanations (SHAP) analysis and representative habitat maps were used for model interpretability. Among the evaluated classifiers, the ExtraTrees model showed the most stable discrimination performance across the training and validation cohorts. The HabitatAll model achieved an AUC of 0.891 in both cohorts, while the Combined model showed the best overall predictive performance, with AUCs of 0.944 and 0.891 in the training and validation cohorts, respectively, and the lowest Brier scores. SHAP analysis identified T1C-derived and H3-related habitat features as major contributors to recurrence prediction. Representative habitat maps showed that H3 was frequently distributed in the peripheral portion of the tumor in both recurrent and non-recurrent cases. This study developed a promising habitat-based multimodal radiomics model for early postoperative recurrence prediction in craniopharyngioma. The Combined model demonstrated the most favorable overall discrimination and calibration. H3-related habitat features may capture intratumoral heterogeneity associated with recurrence risk. However, given the retrospective single-center design, these findings should be interpreted as preliminary and require further validation in larger multicenter cohorts.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s10143-026-04483-8.

Introduction

Craniopharyngiomas are benign (WHO grade I) tumors that arise in the sellar and suprasellar regions and arise from the embryonal epithelium of Rathke’s pouch [1]. Craniopharyngiomas are relatively rare tumors, with an incidence of 0.2 per 100,000 person-years and accounting for 3% of all intracranial tumors [2, 3]. Although non-malignant, CP often impairs important neurological, visual, and endocrine functions due to its proximity to vital structures such as the optic chiasm and hypothalamus. The age of onset showed a bimodal distribution, and it could occur in children (5–15 years old) and adults (45–60 years old) [4, 5]. Among the two main pathological subtypes of adamantinomatous CP (ACP) and papillary CP (PCP), ACP is more common in children [6]. Due to its invasiveness and cystic nature, it has a high recurrence rate and is difficult to operate. Craniopharyngioma recurrence is common, with reported recurrence rates ranging from 9% to 62% [7], and although gross total resection (GTR) is considered to be the most effective method of reducing CP recurrence rates, the tumor recurrence rate after GTR is still relatively high, at approximately 11.1%-25% [8]. And approximately 20%-30% of patients will recur/progress within five years of initial treatment [9].

The application of radiomics in predicting tumor recurrence has garnered increasing attention in recent years [10, 11]. Traditional methods for tumor diagnosis and recurrence prediction—which primarily rely on clinical examination, pathological evaluation, and conventional imaging—often suffer from operational subjectivity and low sensitivity to subtle structural changes within the tumor, exhibiting notable limitations, particularly in the prediction of early recurrence [12–14]. The advancement of radiomics has led to numerous studies aiming to reveal the biological heterogeneity of tumors and predict their prognosis through the quantitative analysis of imaging features. However, early, traditional radiomics, which calculated global features extracted from the Whole Tumor Volume (WTV), presented a key limitation: by treating the tumor as a homogeneous whole, they failed to fully capture its fine spatial heterogeneity [15, 16]. To overcome this, the field has evolved toward subregional Analysis (subregional Radiomics), with Tumor Habitat Analysis being particularly prominent. Habitat analysis is an advanced quantitative approach that utilizes unsupervised clustering techniques to partition the tumor interior into multiple distinct “Habitats,” each characterized by different radiomic signatures and generally considered to represent a unique pathophysiological state within the tumor microenvironment (e.g., hypoxia, necrosis, hyperperfusion). This methodology provides a more robust and biologically interpretable description of tumor characteristics than traditional global features [17–19], offering a more accurate and spatially information-rich quantitative analysis, which is especially well-suited for the precise prediction of tumors at high risk for recurrence.

This study presents an MR-based multi-omics modeling framework integrating imaging and clinical data, with a particular focus on habitat analysis. Three MR sequences—T1, T1C, and T2—were employed to characterize intratumoral subregions through unsupervised clustering of voxel-level features, forming habitat-based representations of tumor biology. We hope to develop an effective prediction model for the early recurrence of craniopharyngioma after surgery based on the above framework, thereby providing effective individualized treatment for patients with craniopharyngioma.

Materials & methods

Study design and population

A retrospective cohort study was conducted on 136 consecutive patients who underwent surgical treatment at the Second Hospital of Lanzhou University between February 2019 and August 2024 and were subsequently confirmed to have craniopharyngioma by postoperative histopathology. This study was conducted in accordance with the Declaration of Helsinki Ethical approval was granted by the Ethics Committee of the Second Hospital of Lanzhou University. Due to the retrospective nature of the study, the requirement for informed consent was waived Clinical data, biochemical indices and neuroimaging were collected.

Patients who met the following criteria were enrolled in this study: (1) primary surgical treatment for craniopharyngioma at our institution; (2) histologically confirmed diagnosis; (3) complete preoperative MRI scans (including T1WI, T2WI, and T1C); and (4) comprehensive medical records. Exclusion criteria included: (1) history of prior treatment (e.g., surgical resection or radiotherapy); (2) severe dysfunction of vital organs, such as the heart, liver, or kidneys; (3) incomplete clinical data; and (4) follow-up durations shorter than the required period.

Our institutional treatment strategy strictly adheres to international consensus. We prioritize gross total resection (GTR) whenever feasible. In cases where complete removal poses an unacceptable risk of severe neurological or endocrinological morbidity, subtotal resection (STR) is performed, followed immediately by adjuvant radiotherapy. Preoperative hypothalamic involvement was assessed using the Puget grading system and was incorporated into surgical planning at our institution. In general, patients with Puget grade 0 disease were managed with an intended gross total resection strategy. Patients with Puget grade 2 disease were prospectively considered at high risk for hypothalamic injury and were therefore generally managed with a planned subtotal resection followed by adjuvant radiotherapy. For patients with Puget grade 1 disease, maximal safe resection was attempted intraoperatively; if gross total resection could not be achieved safely, these patients were managed according to the subtotal resection plus postoperative radiotherapy strategy. Thus, the Puget grade served as an important preoperative anatomical reference for balancing maximal resection against the risk of hypothalamic damage. However, Puget grade itself was not used as an inclusion or exclusion criterion for the present study.

Ultimately, 106 patients were recruited from an initial pool of 136 candidates. As shown in Fig. 1, from the initial 136 surgically treated patients, 30 were excluded because of second surgical treatment (n = 6), incomplete MRI data (n = 21), or poor imaging quality (n = 3), leaving 106 patients for final analysis. These exclusions were based solely on the predefined study criteria and were not related to Puget grade, planned extent of resection, or receipt of adjuvant radiotherapy. The follow-up period extended for at least one year, with data collected up to August 2025. Follow-up assessments occurred at one month and three months postoperatively, and every six months thereafter, with MRI scans performed at each visit. Two radiologists independently reviewed the images to assess recurrence. Recurrence was assessed as a binary clinical outcome based on postoperative MRI follow-up. All patients underwent MRI examination (T1WI, T2WI, and T1C) within 3 days before surgery. Given the retrospective single-center design and the limited number of recurrence events, a randomized hold-out validation strategy was adopted to achieve the best possible baseline comparability between the training and validation cohorts. The dataset was divided into training and validation cohorts at a ratio of 7:3. This randomization process was repeated until most major clinical characteristics were well balanced between the two cohorts. This approach is commonly used in preliminary radiomics studies under limited-sample conditions. To reduce the risk of overfitting, all preprocessing, feature selection, model construction, and parameter optimization procedures were performed strictly within the training cohort, while the validation cohort was reserved for independent performance evaluation. Figure 2 outlines the overall analytical workflow for radiomics model development and evaluation.

Fig. 1.

Patient selection process and cohort allocation

Fig. 2.

Overall analytical workflow for habitat-based radiomics analysis

Image preprocessing and tumor segmentation

ROI Annotation: Tumor regions were manually delineated on cross-sectional T1C images using ITK-SNAP v4.2.0 by two experienced radiologists [20]. Discrepancies were reviewed and resolved by a senior radiologist with over 20 years of experience in diagnostic imaging to ensure annotation accuracy and consistency. The T1 and T2 images were rigidly and affinely registered to the T1C sequence to achieve precise spatial alignment across modalities, enabling the direct application of the T1C-based segmentation mask to all sequences for consistent ROI definition.

Preprocessing Protocol: To standardize spatial resolution among different scanners, all MR images were resampled to an isotropic voxel size of 1 mm³ using interpolation. Intensity normalization was performed to minimize inter-scanner and inter-sequence variability, thereby enhancing feature reproducibility across the dataset.

Habitat subregion modeling

To characterize intratumoral heterogeneity at a fine-grained scale, we established a voxel-level habitat segmentation framework based on quantitative radiomic profiling from MR images. A 5 × 5 × 5 neighborhood kernel was applied across the tumor volume to compute localized intensity and texture descriptors, generating 39 local features for each voxel. This procedure was independently performed on all MR sequences (T1, T1C, and T2), after which the sequence-specific local features were concatenated, resulting in a 117-dimensional voxel-level feature vector. These features encompassed both first-order statistics (e.g., mean, standard deviation, kurtosis) and higher-order texture attributes (e.g., GLCM contrast, GLSZM zone entropy), providing a comprehensive representation of the tumor’s internal architecture (details in Supplementary 1 A). The integrated feature space was subsequently clustered using K-means to delineate candidate subregions, with the optimal cluster number (3–10) determined by the Calinski–Harabasz index to ensure intra-cluster compactness and inter-cluster separability [21]. Representative habitat maps are illustrated in Fig. 3.

Fig. 3.

Presents the habitat regions generated

Feature extraction and selection

Radiomic features were systematically categorized into three primary domains—morphological, intensity-based, and texture-based—to comprehensively capture the phenotypic characteristics of tumor tissues. Morphological features described the lesion’s geometric attributes, including shape, size, and margin irregularity; intensity-based features reflected voxel value distributions; and texture features quantified spatial intensity variations through matrix-based methods such as the Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Size Zone Matrix (GLSZM), and Neighboring Gray Tone Difference Matrix (NGTDM).

Feature extraction was performed at both the global and subregional levels. Global features were derived from the entire tumor volume (VOI), while subregional features were separately extracted from each habitat generated via unsupervised clustering. To ensure spatial consistency and reduce potential noise from border voxels, a K-Nearest Neighbor (KNN)-based relabeling approach was applied to refine the cluster assignments. The value of k was selected by evaluating candidate neighborhood sizes and choosing the one that maximized the Calinski–Harabasz score [22]. All features were computed using the PyRadiomics library (v3.0.1) in accordance with the Imaging Biomarker Standardization Initiative (IBSI) guidelines.

In our experiments, we compared the performance of the HabitatAll model—constructed using habitat features from all MR sequences (T1, T1C, and T2)—with models derived from individual sequences, namely HabitatT1C, HabitatT1, and HabitatT2. To integrate features from multiple regions and sequences, we adopted an early-fusion strategy, concatenating features before model construction to generate a unified and comprehensive feature representation for subsequent learning and prediction.

All extracted features were first standardized using Z-score normalization to ensure comparability across different scales and modalities. A univariate filtering step was then applied using independent t-tests, retaining features significantly associated with the target outcome (p < 0.05). To address multicollinearity, pairwise Pearson correlation analysis was performed, and for each highly correlated pair (r > 0.9), the feature with the lower variance was removed.

After preliminary filtering, the Minimum Redundancy Maximum Relevance (mRMR) algorithm was applied to rank all candidate radiomic features according to their relevance to recurrence and redundancy among features. Based on this ranking, and following the common 1:10 rule of thumb to limit the number of candidate predictors relative to the sample size [23], we retained the top 16 features for downstream modeling to reduce the risk of overfitting. Subsequently, Least Absolute Shrinkage and Selection Operator (LASSO) regression was employed to perform embedded feature selection. This method imposes an L1 penalty to shrink coefficients of less informative features toward zero, effectively retaining only the most predictive subset. The optimal regularization parameter (λ) was determined via 10-fold cross-validation. This multi-stage pipeline produced a compact, stable, and discriminative feature set, ensuring the robustness and interpretability of subsequent models. All preprocessing, feature filtering, feature ranking, and feature selection procedures described above were performed strictly within the training cohort only. No information from the validation cohort was used during these steps. The derived feature-selection pipeline was subsequently applied to the validation cohort for independent evaluation.

Modeling and evaluation of models

Multiple predictive signatures were developed to represent distinct spatial contexts of tumor characterization, each following the same feature selection pipeline and trained using complementary machine learning classifiers to improve model robustness.

Radiomics Signature (Rad): Constructed using global radiomic features extracted from the entire tumor region. Models were trained using Support Vector Machine (SVM) for interpretability, and ensemble classifiers such as Random Forest and ExtraTrees to capture nonlinear relationships. This signature represented global tumor-level imaging characteristics without spatial habitat partitioning and served as a reference model for comparison with habitat-based signatures.

Habitat Signatures: Developed based on habitat-derived subregional features to capture fine-grained intratumoral heterogeneity. Separate models were constructed for each MR sequence—HabitatT1, HabitatT1C, and HabitatT2—as well as a HabitatAll model that integrated all sequences via early feature fusion. Sequence-specific habitat signatures were constructed to evaluate the complementary contributions of different MRI sequences to the characterization of intratumoral heterogeneity. In contrast, the whole-tumor radiomics signature was used as a global imaging reference model. Unlike the whole-tumor radiomics signature, habitat signatures were constructed using radiomic features extracted from spatially defined tumor subregions generated by habitat clustering.

Clinical Signature: Built using available clinical variables with the same feature selection and modeling procedures applied to imaging-based signatures, ensuring methodological consistency across different input data types.

Combined Signature: Significant clinical parameters (p < 0.05) were integrated with the best-performing habitat through feature-level fusion. A nomogram was then constructed to visualize the Combined model and facilitate clinical application.

Candidate classifiers, including Support Vector Machine (SVM), Random Forest (RF), and Extra Trees (ET), were evaluated to determine the optimal classifier framework. The selected classifier was subsequently applied consistently across all predictive signatures, including clinical, whole-tumor radiomics, habitat-based, and Combined models, to ensure fair comparison among different feature representations. Hyperparameter optimization and feature selection were performed within the training cohort. The validation cohort was not involved in feature selection, classifier tuning, or model optimization. Instead, it was reserved for comparative assessment of candidate classifiers and subsequent performance evaluation.

Model performance was comprehensively evaluated using Receiver Operating Characteristic (ROC) analysis, with the Area Under the Curve (AUC) as the primary metric, supplemented by accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Model calibration was assessed using calibration plots, the Hosmer–Lemeshow goodness-of-fit test, calibration intercept, calibration slope, and Brier score to evaluate the agreement between predicted probabilities and observed outcomes. The clinical utility and net benefit of the models were further quantified using Decision Curve Analysis (DCA) across a range of decision thresholds.

Candidate classifiers were compared according to both discrimination performance and stability across the training and validation cohorts. In addition to AUC magnitude, consistency of performance between cohorts was considered as an indicator of model robustness, with the aim of reducing the likelihood of selecting an overfitted classifier based solely on training performance. The performance of all candidate classifiers was reported transparently. The classifier showing the most favorable balance between discrimination and stability across cohorts was ultimately selected for subsequent signature comparison and model presentation.

Statistical analysis

We conducted descriptive and inferential statistical analyses using Python 3.7.12 on the OnekeyAI platform (v3.1.8). Categorical variables were analyzed using the chi-square test, while continuous variables were compared using the independent samples t-test. Feature extraction was conducted with PyRadiomics v3.0.1, and statistical modeling was implemented using statsmodels v0.13.2. All machine learning algorithms, including SVM and tree-based classifiers, were developed using scikit-learn v1.0.2.

Results

Clinical characteristics

A total of 106 patients (55 males and 51 females; 82 in the non-recurrent group and 24 in the recurrent group) were included in this study. The mean age of the overall cohort was 41.05 ± 18.95 years. Table 1 summarizes the baseline clinical characteristics of the study population, including extent of resection and Puget grade distribution.

Table 1.

Baseline characteristic of the study population

Feature name	ALL	Train	Validation	P-value
Weight	60.02 ± 19.06	60.03 ± 20.05	60.00 ± 16.84	0.993
Height	159.29 ± 17.63	158.36 ± 18.34	161.44 ± 15.91	0.546
Age	41.05 ± 18.95	40.89 ± 19.19	41.41 ± 18.70	0.866
Total duration of hospitalization	30.61 ± 19.99	31.46 ± 19.82	28.66 ± 20.54	0.349
Blood loss	260.09 ± 279.32	255.00 ± 312.11	271.88 ± 186.19	0.223
Operation duration	4.50 ± 1.38	4.39 ± 1.20	4.75 ± 1.73	0.431
K	3.91 ± 0.33	3.88 ± 0.34	3.98 ± 0.29	0.194
Na	138.84 ± 3.86	138.62 ± 3.07	139.34 ± 5.28	0.246
Cl	103.22 ± 3.99	103.16 ± 3.40	103.35 ± 5.16	0.513
ALB	42.48 ± 3.32	42.21 ± 3.37	43.12 ± 3.16	0.199
WBC	6.42 ± 1.99	6.36 ± 2.02	6.54 ± 1.96	0.635
Gender				0.038
Female	51(48.11)	41(55.41)	10(31.25)
Male	55(51.89)	33(44.59)	22(68.75)
Polyuria				0.692
No	90(84.91)	64(86.49)	26(81.25)
Yes	16(15.09)	10(13.51)	6(18.75)
Headache				0.61
No	42(39.62)	31(41.89)	11(34.38)
Yes	64(60.38)	43(58.11)	21(65.62)
Decreased visual acuity				0.21
No	34(32.08)	27(36.49)	7(21.88)
Yes	72(67.92)	47(63.51)	25(78.12)
Visual field defect				1.0
No	66(62.26)	46(62.16)	20(62.50)
Yes	40(37.74)	28(37.84)	12(37.50)
Suprasellar region				0.803
No	20(18.87)	13(17.57)	7(21.88)
Yes	86(81.13)	61(82.43)	25(78.12)
Sellar region				1.0
No	58(54.72)	40(54.05)	18(56.25)
Yes	48(45.28)	34(45.95)	14(43.75)
Calcification				0.472
No	59(55.66)	39(52.70)	20(62.50)
Yes	47(44.34)	35(47.30)	12(37.50)
GTR				0.801
No	56(52.83)	38(51.35)	18(56.25)
Yes	50(47.17)	36(48.65)	14(43.75)
Histopathology				1.0
ACP	74(69.81)	52(70.27)	22(68.75)
PCP	32(30.19)	22(29.73)	10(31.25)
Puget				0.847
0	16(15.1)	12(16.2)	4(12.5)
1	63(59.4)	44(59.5)	19(59.4)
2	27(25.5)	18(24.3)	9(28.1)
CSF leak				0.144
No	75(70.75)	56(75.68)	19(59.38)
Yes	31(29.25)	18(24.32)	13(40.62)

Abbreviations: K potassium, Na sodium, Cl chloride, ALB albumin, WBC white blood cell count, CSF cerebrospinal fluid

A comprehensive univariate analysis was conducted for all clinical variables to evaluate their individual associations with the outcome. Odds ratios (ORs) and corresponding p-values were calculated for each variable, and those with p < 0.05 were considered statistically significant. These significant predictors were subsequently incorporated into a multivariable logistic regression model to identify independent clinical factors contributing to the outcome. These significant predictors were subsequently incorporated into a multivariable logistic regression model to evaluate their independent associations with recurrence. The multivariate analysis revealed that gender was significantly associated with the outcome, with an odds ratio (OR) of 1.273, a 95% confidence interval (CI) of 1.085 to 1.493, and a P-value of less than 0.05 (Tables 2, 3, 4 and 5).

Table 2.

Univariable Analysis of clinical features

Feature name	OR	95% CI	P-value
Diabetes	0.787	0.5190–1.1940	0.341
K	0.896	0.7050–1.1400	0.45
Histopathology	0.934	0.7800–1.1190	0.53
EEA	0.955	0.8040–1.1330	0.653
Operation duration	0.966	0.9010–1.0350	0.398
Polyuria	0.966	0.7590–1.2300	0.813
ALB	0.984	0.9600–1.0080	0.277
GTR	0.986	0.8350–1.1630	0.883
Na	0.989	0.9630–1.0150	0.479
CSF leak	0.990	0.8170–1.2010	0.932
WBC	0.991	0.9510–1.0330	0.717
Weight	0.997	0.9930–1.0010	0.27
Age	0.997	0.9920–1.0010	0.206
Height	0.997	0.9930–1.0020	0.286
Blood loss	1.000	1.0000–1.0000	0.99
Total duration of hospitalization	1.002	0.9970–1.0060	0.56
Cl	1.005	0.9810–1.0300	0.724
Headache	1.007	0.8510–1.1900	0.947
Decreased visual acuity	1.012	0.8520–1.2010	0.909
Visual field defect	1.033	0.8710–1.2250	0.75
Craniotomy	1.048	0.8820–1.2440	0.653
Calcification	1.053	0.8930–1.2420	0.601
Sellar region	1.067	0.9040–1.2590	0.516
Suprasellar region	1.096	0.8830–1.3610	0.481
Hypertension	1.119	0.8280–1.5140	0.536
Gender	1.273	1.0850–1.4930	< 0.05

Abbreviations: EEA endoscopic endonasal approach

Table 3.

Model performance of different machine learning algorithms in each cohort

Cohort	Model name	Accuracy	AUC	95% CI	Sensitivity	Specificity	PPV	NPV
Train	SVM	0.959	0.980	0.954–1.000	0.941	0.965	0.889	0.982
	RandomForest	0.905	0.948	0.893–1.000	0.882	0.912	0.750	0.963
	ExtraTrees	0.797	0.891	0.818–0.965	0.882	0.772	0.536	0.957
Validation	SVM	0.906	0.811	0.591–1.000	0.571	1.000	1.000	0.893
	RandomForest	0.844	0.811	0.590–1.000	0.714	0.880	0.625	0.917
	ExtraTrees	0.812	0.889	0.775–1.000	1.000	0.760	0.538	1.000

Table 4.

Prediction performance of habitat based rad signatures

Cohort	Signature	Accuracy	AUC	95% CI	Sensitivity	Specificity	PPV	NPV
Train	Clinical	0.730	0.848	0.7466–0.9489	0.882	0.684	0.455	0.951
	Radiomics	0.811	0.878	0.7792–0.9773	0.882	0.789	0.556	0.957
	HabitatT2	0.811	0.896	0.8234–0.9681	0.882	0.789	0.556	0.957
	HabitatT1	0.622	0.728	0.6257–0.8294	0.941	0.526	0.372	0.968
	HabitatT1C	0.703	0.813	0.7156–0.9098	0.882	0.649	0.429	0.949
	HabitatAll	0.797	0.891	0.8176–0.9646	0.882	0.772	0.536	0.957
	Combined	0.838	0.944	0.8959–0.9926	1.000	0.789	0.586	1.000
Validation	Clinical	0.531	0.731	0.5447–0.9182	1.000	0.400	0.318	1.000
	Radiomics	0.781	0.823	0.6653–0.9804	0.857	0.760	0.500	0.950
	HabitatT2	0.719	0.760	0.5896–0.9304	1.000	0.640	0.437	1.000
	HabitatT1	0.625	0.594	0.3592–0.8293	0.571	0.640	0.308	0.842
	HabitatT1C	0.594	0.709	0.5318–0.8854	1.000	0.480	0.350	1.000
	HabitatAll	0.812	0.889	0.7747–1.0000	1.000	0.760	0.538	1.000
	Combined	0.844	0.891	0.7794–1.0000	1.000	0.800	0.583	1.000

Table 5.

Calibration performance of candidate models in the training and validation cohorts

Cohort	Signature	Calibration intercept	Calibration slope	Brier score
Train	Clinical	1.832	2.819	0.129
	Radiomics	-2.994	2.618	0.253
	HabitatT2	-1.713	2.639	0.190
	HabitatT1	-0.780	0.174	0.252
	HabitatT1C	-1.228	2.153	0.189
	HabitatAll	-1.331	4.828	0.184
	Combined	0.041	1.135	0.084
Validation	Clinical	-0.297	0.980	0.160
	Radiomics	-2.337	1.919	0.277
	HabitatT2	-1.141	1.321	0.211
	HabitatT1	-1.039	0.020	0.288
	HabitatT1C	-0.970	0.783	0.200
	HabitatAll	-1.357	5.545	0.202
	Combined	0.284	1.367	0.100

Habitat signature

In this study, a total of 1,561 handcrafted radiomic features were extracted from each region of interest (ROI) for every MR sequence. These features encompassed three principal categories: 14 shape descriptors describing tumor morphology, 306 first-order intensity features reflecting voxel-wise intensity statistics, and 1,241 texture features quantifying spatial patterns through GLCM, GLRLM, GLSZM, and NGTDM matrices.

For the Habitat models, radiomic features were extracted separately from three spatially distinct intratumoral subregions generated via unsupervised clustering. Each MR sequence (T1, T1C, and T2) contributed 1,561 features per region, resulting in 4,683 features (3 regions × 1,561 features) for each sequence-specific habitat model (HabitatT1, HabitatT1C, HabitatT2). The HabitatAll model, which integrated all three sequences and three regions, contained 14,049 features (3 sequences × 3 regions × 1,561 features) after feature-level fusion (Fig. 4).

Fig. 4.

Distribution of the extracted radiomics features. (a) Pie chart showing the percentage of each feature group. (b) Bar chart showing the absolute count of features in each corresponding feature group

To prevent overfitting and retain the most informative predictors, feature selection was performed using Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression with 10-fold cross-validation. The optimal regularization parameter (λ) was determined by minimizing the Mean Squared Error (MSE), ensuring a balance between model sparsity and stability. The coefficient paths and MSE curve are shown in Fig. 5, along with a histogram of selected features related to intratumoral heterogeneity.

Fig. 5.

Feature selection using least absolute shrinkage and selection operator (LASSO) regression. (a) LASSO coefficient trajectories of candidate radiomic features. (b) Cross-validation curve for selection of the optimal penalty parameter (λ). The vertical dashed line indicates the optimal λ value. (c) Radiomic features retained by LASSO regression and their corresponding coefficients. Feature names follow the standard radiomics nomenclature, where wavelet and local binary pattern (LBP) indicate image transformation methods, GLCM represents gray-level co-occurrence matrix, GLSZM represents gray-level size zone matrix, and h1/h2/h3 indicate habitat subregions

Different number of clusters

We systematically evaluated clustering performance by varying the number of clusters between 3 and 10 and determining the optimal configuration using the Calinski–Harabasz (CH) index, which measures inter-cluster separation and intra-cluster compactness. The CH score reached its maximum when the number of clusters was set to three, indicating that this configuration provided the most stable and well-differentiated subregional segmentation (Figs. 6 and 7).

Fig. 6.

Identification of intratumoral habitats using unsupervised clustering. (a) Calinski–Harabasz index analysis for determining the optimal number of clusters. (b) Three-dimensional visualization of voxel-level feature distribution after dimensionality reduction. Each point represents an individual voxel, and different colors indicate distinct habitat clusters identified by k-means clustering

Fig. 7.

ROC Curves of different models in each cohorts

Evaluation metrics

Across the training and validation cohorts, all models demonstrated satisfactory classification performance. In the training cohort, the SVM model achieved the highest AUC (0.980) and accuracy (0.959), with sensitivity and specificity of 0.941 and 0.965, respectively. However, its generalization performance decreased notably in the validation cohort (AUC = 0.811, accuracy = 0.906). The Random Forest model exhibited stable yet moderate performance (AUC = 0.948 in training and 0.811 in validation). The ExtraTrees model achieved an AUC of 0.891 in the training cohort and 0.889 in the validation cohort, maintaining excellent sensitivity (1.000) and a balanced trade-off between specificity (0.760) and accuracy (0.812) during validation evaluation.

Among the compared classifiers, the ExtraTrees model demonstrated the most stable and generalizable performance, showing consistent AUC values between training and validation cohorts with minimal overfitting. Its superior sensitivity and robust predictive stability suggest that it provides the most reliable performance for this classification task.

SHAP value

SHAP (SHapley Additive exPlanations) analysis was conducted to enhance interpretability of the ExtraTrees-based Habitat model by quantifying the contribution of each feature to the model’s output. SHAP values provide both instance-level interpretability and global feature importance, offering insights into the underlying decision-making process of the model (Fig. 8).

Fig. 8.

(a) SHAP summary plot demonstrating the distribution of feature contributions across all samples. Features derived from the T1C sequence appear prominently, contributing two of the top three most influential predictors. (b) Mean SHAP value bar plot highlighting global feature importance. Features originating from habitat region H3 constitute three of the highest-ranked predictors, including two within the top tier, underscoring the strong discriminative relevance of this subregion. (c) Representative SHAP force plot illustrating instance-level interpretability and the directional influence of key features on the model output

In the SHAP-based feature importance analysis, the T1C sequence demonstrated notable dominance, contributing two of the top three most influential features, underscoring its strong predictive relevance among all MR modalities. From a spatial perspective, habitat region H3 emerged as particularly significant, providing three of the selected key features, including two within the top-ranked group. These results indicate that both the T1C sequence and the H3 subregion play crucial roles in driving model predictions, suggesting that this specific imaging modality and spatial zone may capture biologically meaningful characteristics most closely associated with the clinical outcome.

To further enhance spatial interpretability, representative habitat maps from both non-recurrent and recurrent cases were reviewed. In these illustrative cases, the H3 habitat was predominantly distributed in the peripheral portion of the tumor, providing visual support for the habitat-based spatial patterns underlying the SHAP findings (Fig. 9).

Fig. 9.

Representative habitat maps in non-recurrent and recurrent cases. Panels a and b show non-recurrent cases, and panels c and d show recurrent cases. In these representative images, the H3 habitat was predominantly located in the peripheral portion of the tumor, providing visual support for the spatial interpretation of the habitat-based model

Signature comparsion

Calibration was assessed using calibration plots, the Hosmer–Lemeshow (HL) goodness-of-fit test, calibration intercept, calibration slope, and Brier score. The HL test showed nonsignificant results in both the training and validation cohorts (training: 0.904; validation: 0.482; both p > 0.05), suggesting no statistically significant lack of fit. However, calibration performance varied across candidate models when assessed using additional quantitative metrics. Overall, the Combined model showed the most favorable calibration profile, with a calibration intercept of 0.041, calibration slope of 1.135, and Brier score of 0.084 in the training cohort, and corresponding values of 0.284, 1.367, and 0.100 in the validation cohort. The Clinical model showed relatively well-aligned calibration intercept and slope in the validation cohort (-0.297 and 0.980, respectively), but its Brier score (0.160) remained higher than that of the Combined model. In contrast, several single-signature models showed less satisfactory calibration performance. Notably, the HabitatAll model demonstrated marked deviation in calibration slope in both the training and validation cohorts (4.828 and 5.545, respectively), despite moderate Brier scores (0.184 and 0.202), while the HabitatT1 model showed very low calibration slopes (0.174 and 0.020, respectively) together with relatively high Brier scores (0.252 and 0.288). These findings suggest that, among the evaluated models, the Combined model provided the most favorable probability calibration, whereas calibration quality varied substantially across the remaining candidate models.

Delong test

The Delong test was used to compare the AUCs among different models. In both the training and validation cohorts, the Habitat and Combined models achieved significantly higher AUCs than most other models (p < 0.05), confirming that their superior predictive performance was statistically significant (Figs. 10, 11 and 12).

Fig. 10.

ROC curves of all cohorts

Fig. 11.

Calibration curve on all cohort

Fig. 12.

Delong test of different signature

Clinical use

Decision Curve Analysis (DCA)

Figure 13 displays the DCA curves for both the training and validation cohorts. The analysis of these curves indicates that our Combined model offers a notable advantage in terms of net benefit derived from predicted probabilities.

Fig. 13.

Different signatures’ decision curve on all cohort

Nomogram: We employed a Nomogram to visualize the results of our Combined model (Fig. 14).

Fig. 14.

shows the nomogram for clinical use

Discussion

The recurrence rate of craniopharyngiomas after surgery is high, even after gross total resection [6, 24]. If the tumor is left behind during surgery to protect critical structures, the postoperative recurrence rate can reach as high as 50%. Although postoperative radiotherapy can effectively delay tumor recurrence [25], it is typically reserved for patients with multiple recurrences or those who are older and unable to tolerate surgery. Given the proximity of the tumor to critical structures such as the hypothalamus, optic nerves, and major blood vessels, identifying patients at high risk for postoperative recurrence is crucial [26]. Therefore, we conducted this study with the goal of implementing screening measures for individuals at high risk of early recurrence, which could facilitate the development of personalized treatment strategies and postoperative monitoring plans, ultimately improving patient survival rates and quality of life.

Previous imaging radiomic studies on craniopharyngiomas have largely focused on differentiating craniopharyngiomas from other tumors or predicting tumor pathology before surgery [27–30]. Huang et al. developed a non-invasive preoperative risk stratification model, which identified high expression of fibroblast activation protein (FAP) as being associated with recurrent adamantinomatous craniopharyngioma (ACP) [31]. Their model, which integrated T1C and T2 sequence features, demonstrated robust performance in predicting FAP expression, with an internal validation AUC of 0.822 and an external validation AUC of 0.686. Our study, however, focuses more on the tumor heterogeneity of craniopharyngiomas by subdividing the tumor into distinct habitat subregions and constructing predictive models using radiomic features from these subregions. Unlike previous studies that mainly focused on preoperative diagnosis, pathological classification, or molecular characterization, our study aimed to predict postoperative recurrence risk by integrating habitat-based radiomic features with clinical information. Studies have shown that in ACP, the senescence-associated secretory phenotype (SASP) acts on surrounding epithelial tumor cells, promoting tumor proliferation and invasion while also fostering the development of the tumor microenvironment [32–34]. Habitat analysis can differentiate various tumor cells and their microenvironment and classify them into specific regions, thereby revealing the spatial heterogeneity of the tumor [35, 36]. Similar habitat-based radiomics approaches have been investigated in other brain tumors, supporting the potential of spatial heterogeneity analysis for prognostic modeling. However, its application in craniopharyngioma recurrence prediction remains largely unexplored.

In our study, we extracted features from three different imaging sequences to obtain a more comprehensive and accurate set of radiomic characteristics. We then performed clustering analysis on these features, with each sequence resulting in different H1, H2, and H3 subregions. After feature selection, we identified the ExtraTrees model as the classifier. Based on these results, we developed several predictive models for comparison, and we found that models incorporating habitat analysis outperformed both clinical prediction models and traditional imaging genomics models. The models showing the best performance were the HabitatAll and Combined models, which achieved AUCs of 0.812 and 0.844, respectively, in the validation cohort. Additional calibration analysis further showed that calibration performance varied considerably across candidate models. Among all evaluated signatures, the Combined model demonstrated the most favorable overall calibration, with the lowest Brier scores in both the training and validation cohorts and calibration intercept and slope values relatively close to their ideal values (0 and 1, respectively). Although the Clinical model exhibited a comparatively well-behaved calibration intercept and slope in the validation cohort, its overall Brier score remained higher than that of the Combined model, indicating inferior overall probability accuracy. By contrast, several single-signature models showed suboptimal calibration characteristics. In particular, the HabitatAll model displayed markedly inflated calibration slopes, and the HabitatT1 model showed extremely low slopes, suggesting that the predicted probabilities generated by these models were less well calibrated despite some discriminatory ability. Taken together, these results indicate that the Combined model not only achieved encouraging discrimination but also provided more reliable probability estimates than most alternative models, supporting its use as the primary model for presentation and clinical interpretation.

We applied SHAP analysis to assess the contribution of each feature to individual predictions, thereby improving the interpretability of the machine learning model [37]. In the present study, SHAP analysis showed that features derived from the T1C sequence had high overall importance in recurrence prediction. At the same time, several of the top-ranked features were derived from the T1 sequence. This pattern may reflect the complementary value of contrast-enhanced and non-enhanced imaging. The T1C sequence can more clearly depict tumor margins, vascularity, and necrotic or heterogeneous enhancing components, thereby providing important information on macroscopic characteristics such as border definition and angiogenesis-related heterogeneity [38]. By contrast, T1 images may better capture relatively stable internal structural features of the tumor, including tissue composition and subtle texture variation, which may be less affected by contrast administration and therefore potentially more reproducible in radiomic analysis [39]. This interpretation is also consistent with the findings of McHugh et al. (2021), who reported that although contrast-enhanced imaging may provide richer biological information, non-enhanced sequences may offer advantages in radiomic feature repeatability and stability [40].

Beyond feature ranking, qualitative inspection of the habitat maps provided additional spatial insights. In relatively small tumors, the H1 region often occupied most of the lesion, whereas H2 and H3 were less conspicuous. As tumor volume increased, the proportion of H1 appeared to decrease, while H2 and H3 became more spatially prominent. Although this observation was qualitative rather than quantitatively measured, it suggests that these habitats may reflect dynamic intratumoral heterogeneity that evolves during tumor growth, with H2 and H3 potentially representing more complex microenvironmental states emerging as the lesion enlarges. Representative habitat maps further showed that the H3 region was frequently distributed in the peripheral portion of the tumor in both recurrent and non-recurrent cases. This suggests that peripheral localization may represent a general spatial characteristic of the H3 habitat rather than a recurrence-specific pattern. Nevertheless, SHAP analysis indicated that H3-derived features ranked among the most influential predictors, suggesting that the prognostic relevance of H3 may be more closely related to its radiomic characteristics than to peripheral location alone.

We found that the H3 region was frequently located in the peripheral portion of the tumor, and H3-derived features were strongly associated with recurrence prediction. Previous studies in other tumor types have suggested that peripheral tumor regions may contain heterogeneous imaging characteristics associated with biological variation. However, whether similar biological mechanisms exist in craniopharyngioma remains unclear [41]. Previous studies have shown that cells in the tumor margin or peripheral region are usually in a hypermetabolic state and are closely related to tumor recurrence and invasion [42]. In the study of brain tumors such as gliomas, cell populations at the outer edge of the tumor tend to exhibit more aggressive behavior and easily expand into the surrounding normal tissues [43, 44]. Thus, the H3 region may reflect a spatially distinct peripheral habitat, while its prognostic relevance may be more closely related to the radiomic characteristics captured within this habitat than to peripheral location alone. By partitioning tumors into multiple subregions (e.g., H1, H2, H3), we can reveal the spatial heterogeneity of tumors and provide more accurate prediction of recurrence risk. Similar subregion analysis has been validated in many studies, and the heterogeneity of the tumor margin has a direct relationship with recurrence [45]. This result supports the important biological significance of the peripheral region of the tumor as a region of high risk for recurrence. However, it is important to note that our current habitat identification relies on unsupervised machine learning and early multi-omics feature fusion, which clusters imaging similarities in a high-dimensional space. While this method effectively identifies high-risk biological textures within the H3 region, the mathematical process involves deconstructing original spatial voxel coordinates. As a result, we cannot currently map these abstracted microenvironments back to precise anatomical boundaries, such as the specific walls of the sella turcica, to calculate frequency distributions of microscopic invasion. This technological barrier remains a challenge in translating high-dimensional radiomic clusters into exact anatomical correlates.

In this study, the radiomics analysis of multimodal MRI images combined with habitat analysis was used to develop a model that can predict tumor recurrence. Compared with the traditional overall tumor feature analysis, the subregion analysis can more accurately capture the spatial heterogeneity within the tumor, especially the outer edge of the tumor. This method has been widely used in the prediction and prognostic evaluation of glioma, lung cancer and other tumors, and has been shown to significantly improve the accuracy of prediction [46–48]. The heterogeneity of the outer tumor margin region is particularly important in patients with a higher risk of recurrence, and therefore, it should be analyzed and monitored in detail as a key concern in practical clinical practice. In addition, multimodal models such as HabitatAll model can provide a more comprehensive description of tumor characteristics by integrating radiomic features of T1, T1C and T2 sequences, thereby improving the accuracy of recurrence prediction. This multimodal analysis method can not only improve the accuracy of diagnosis and prognosis, but also provide a reliable basis for the formulation of individualized treatment plans.

In neurosurgical practice, postoperative recurrence is a critical factor affecting the prognosis of brain tumor patients, especially in cases like craniopharyngiomas, which are benign tumors. Although these tumors are of low malignancy, their proximity to critical structures such as the pituitary gland, optic nerves, and other vital brain areas still poses a significant risk of recurrence. Traditional surgical approaches focus on maximizing tumor resection, but complete removal in areas close to these vital structures may lead to unacceptable neurological deficits, particularly affecting vision, endocrine function, and cognition. Therefore, early prediction of recurrence is crucial for neurosurgeons to develop personalized treatment plans and surgical strategies.

The multimodal radiomics model proposed in this study, combined with subregion analysis and the nomogram, provides a valuable clinical tool for early prediction of tumor recurrence. In neurosurgical procedures, understanding high-risk regions for recurrence helps surgeons make more precise preoperative assessments, particularly regarding the heterogeneity of the tumor’s peripheral regions. For example, the H3 region, which is closely associated with recurrence, should be a focal point during surgery, as it may require particular attention to ensure adequate resection while preserving surrounding normal tissue. Preoperative imaging analysis, coupled with recurrence risk prediction, allows for more accurate delineation of tumor boundaries, helping surgeons make optimal decisions for resection while ensuring the safety of neurological function. Furthermore, the application of the nomogram provides guidance for postoperative management and follow-up. By accurately assessing the risk of recurrence, neurosurgeons can decide whether more frequent follow-up or adjunctive treatments are necessary, enabling early detection and intervention for potential recurrence. This multidimensional approach, integrating imaging, clinical data, and surgical strategy, enables neurosurgeons to provide individualized treatment and postoperative management, ultimately improving patients’ quality of life and treatment outcomes. However, future studies integrating intratumoral habitats with anatomical information from surrounding structures may provide a more comprehensive understanding of the relationship between tumor heterogeneity, surgical complexity, and recurrence risk, further improving the clinical applicability of radiomics-based models.

Our study has several limitations. First, this was a retrospective single-center study with a relatively small sample size and a limited number of recurrence events, which may have introduced potential selection bias and limited the generalizability of the model. Although we used an independent training–validation split and repeated random partitioning to improve baseline comparability between cohorts, more rigorous validation strategies, such as nested cross-validation, bootstrap-based optimism correction, temporal validation, or external multicenter validation, may provide a more robust assessment of model generalizability. In addition, because validation-cohort performance contributed to comparative assessment of candidate classifiers, the strict independence of the validation cohort for final model determination was reduced, which may also have introduced potential selection bias. Therefore, the current findings should be interpreted as preliminary, and future studies should include larger independent cohorts through multicenter collaboration, ideally with predefined classifiers or nested resampling strategies. In addition, the extent of resection and use of adjuvant radiotherapy were determined clinically rather than randomly, which should be considered when interpreting recurrence risk. Second, we used T1C as the reference for registration, and differences in spatial resolution and potential distortions among MRI sequences may have affected registration accuracy. Third, the clinical data collected in this study were incomplete, particularly with respect to endocrine-related variables in patients with craniopharyngioma. Future studies should include standardized blood sample collection at fixed time points to harmonize hormone level measurements [49–51]. Fourth, this study mainly focused on intratumoral habitat-based radiomic features and did not quantitatively incorporate surrounding anatomical structures or peritumoral regions. Although clinical factors such as Puget grade and extent of resection partially reflect the relationship between tumors and critical structures, future studies integrating anatomical segmentation of adjacent structures and peritumoral imaging features may provide additional information regarding surgical complexity and recurrence risk. Finally, computed tomography(CT)-derived imaging features were not included in the current analysis. Although MRI provides superior soft-tissue characterization for tumor assessment, CT may offer complementary structural information, and integrating both imaging modalities may allow a more comprehensive radiomic characterization of the tumor.

Additionally, a technical challenge remains in the spatial mapping of high-dimensional radiomic clusters back to precise anatomical boundaries, such as the sella turcica walls. Due to the retrospective nature of this study, we could not achieve a one-to-one microscopic correlation between computational habitats and intraoperative findings of tumor adherence. Future prospective research utilizing intraoperative neuronavigation will be essential to bridge this gap and translate these predictive maps into real-time surgical guidance.

Conclusion

This study developed a promising habitat-based multimodal radiomics model for predicting early postoperative recurrence of craniopharyngioma by integrating multi-sequence MRI features and clinical variables to characterize intratumoral heterogeneity. The Combined model demonstrated the most favorable overall predictive performance among the evaluated signatures. SHAP analysis further suggested that the T1C sequence and H3-related habitat features were important contributors to recurrence prediction. However, given the retrospective single-center design, the current findings should be interpreted as preliminary. Further validation in larger multicenter cohorts, as well as prospective studies incorporating intraoperative neuronavigation, will be necessary to confirm the model’s generalizability and to further explore the anatomical and clinical relevance of high-risk imaging habitats. Such efforts may ultimately contribute to more individualized treatment planning and more precise surgical decision-making.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Kunlin Hou: Conceptualization; Data curation; Imaging preprocessing; Radiomic feature extraction; Machine-learning modeling; Statistical analysis; Visualization; Drafting of the original manuscript.Tengyun Guo: Data curation; Clinical data verification; Assistance with feature extraction; Support in visualization and figure preparation; Manuscript editing.Yadong Wang: Data collection; MRI quality assessment; Assistance with modeling validation; Manuscript review and minor revisions.Jiayu Song: Methodology support; Assistance with radiomic analysis; Model performance evaluation; Manuscript editing and technical review.Qiang Li: Study supervision; Methodological guidance; Critical revision of the manuscript for important intellectual content; Oversight of clinical interpretation; Corresponding author responsibilities.Yawen Pan: Study design; Project administration; Supervision of analysis workflow; Critical manuscript review and editing; Final approval of the version to be submitted; Corresponding author responsibilities.All authors read and approved the final manuscript.

Funding

The authors received no specific funding for this work.

Data availability

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

Declarations

Consent for publication

Not applicable. This study does not include any individual identifiable data.

Ethics approval

This study was approved by the Ethics Committee of the Second Hospital of Lanzhou University (Approval No. 2025 A-534) and conducted in accordance with the Declaration of Helsinki.

Consent to participate

Not applicable. Due to the retrospective nature of the study and the use of anonymized patient data, the requirement for informed consent was waived by the Ethics Committee of the Second Hospital of Lanzhou University.

Competing interests

The authors declare no competing interests.

Clinical trial registration

Not applicable. This study is a retrospective observational cohort study and does not involve any prospective interventional procedures; therefore, clinical trial registration is not required.