Multimodal MRI habitat atlas for personalised re-haemorrhage risk prediction in brainstem cavernous malformations

Abstract

Objective

To develop an MRI habitat atlas that is robust across different haemorrhage stages for haemorrhagic brainstem cavernous malformations (BSCMs) and to integrate habitat-derived imaging biomarkers with clinical variables for precise prediction of symptomatic re-haemorrhage.

Methods

A retrospective cohort of 205 eligible patients from Huashan Main Campus (2015–2020) was randomly divided into a discovery set (n=147) for model development and an internal validation set (n=58). External validation was performed using a prospective temporal cohort from Huashan West Campus (2021–2023, n=94). Each lesion was segmented into six biologically distinct habitats via unsupervised clustering of multiparametric MRI. Habitat-derived radiomic features were combined with clinical variables into a multimodal logistic regression model, which was rigorously validated both internally and temporally.

Results

The 2-year cumulative re-haemorrhage rate was 81% in high-risk lesions versus 26% in low-risk lesions. The combined habitat–clinical model achieved an area under the curve (AUC) of 0.833 (95% CI 0.777 to 0.889) in the training set and 0.851 in the external validation set, significantly outperforming both imaging-only (∆AUC+0.07, p=0.008) and clinical-only models (∆AUC+0.17, p<0.001). The model demonstrated excellent calibration and provided net clinical benefit across decision thresholds of 5%–75%.

Conclusion

The multimodal MRI habitat atlas provides a robust and reproducible tool for individualised re-haemorrhage risk stratification in BSCMs, facilitating clinical decision-making for surveillance or intervention while reducing the likelihood of overtreatment.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Current risk stratification for brainstem cavernous malformations (BSCMs) is hindered by the temporal instability of haemoglobin degradation signatures on MRI, which complicates the identification of patients at high risk for symptomatic re-haemorrhage.

WHAT THIS STUDY ADDS

• This study establishes a novel, stage-independent MRI habitat atlas for BSCMs from a large multi-campus cohort. By capturing intralesional microstructural features that are robust to haemorrhagic fluctuations, it accurately quantifies the 2-year risk of re-haemorrhage.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• This objective prediction tool facilitates personalised clinical decision-making by distinguishing high-risk BSCMs requiring intervention from stable ones, thereby standardising management strategies and optimising patient selection to prevent surgical overtreatment.

Introduction

The prevalence of cerebral cavernous malformations (CCMs) detected by MRI ranges from 0.16% to 0.8% in the general population.^1–3 Lesions located in the brainstem, accounting for 18%–35% of all CCMs, are particularly prone to symptomatic haemorrhage due to the dense concentration of critical neural structures within a confined anatomical volume.^{4 5} Reported annual re-haemorrhage rates for brainstem cavernous malformations (BSCMs) are substantial, with prospective cohort studies indicating a risk of 26.3%–42.4% within the first few years after an initial bleed, compared with 5%–9% per year for previously unruptured lesions.^{4 6 7} Although the traditional view emphasises poor functional outcomes, recent large-scale prospective investigations have observed that a significant proportion of patients can maintain stable neurological status over time, with disease-specific mortality below 2%.^{4 5} This heterogeneity in natural history underscores the critical need for precise risk stratification to tailor individualised management strategies, avoiding overtreatment in low-risk cases while prioritising intervention for high-risk patients.

The strongest established predictors of future haemorrhage in BSCMs are a prior symptomatic bleed and specific qualitative MRI phenotypes, such as the Zabramski I classification.⁸ Emerging molecular evidence reveals that, beyond the well-established CCM1–3 genes, somatic mutations in PIK3CA and MAP3K3 are also implicated in CCM pathogenesis, with the MAP3K3 c.1323C>G (p.Ile441Met) mutation being overwhelmingly associated with the acute haemorrhagic (Zabramski type I) and classic ‘popcorn’ (type II) appearances on MRI.^9–12 Accurate risk stratification is still critically hindered by the inherent limitations of qualitative MRI assessments, which exhibit marked inter-observer variability and limited generalisability.^13–15 Moreover, the dynamic temporal evolution of haemoglobin degradation products confounds conventional imaging analysis, obscuring the underlying vascular architecture and rendering radiomic features unstable over time.^{16 17}

Conventional MRI-based radiomics models for predicting re-haemorrhage in BSCMs face significant challenges, including temporal instability caused by the dynamic evolution of haemoglobin degradation products, conflation of the lesions with surrounding haematoma in whole-lesion analyses and susceptibility to overfitting, especially with moderate sample sizes. In response, the ‘tumour habitat’ framework, originally developed in oncology, has emerged as a promising alternative.^18–20 This approach employs unsupervised clustering of multiparametric MRI to partition tumours into biologically interpretable and temporally stable subregions, capturing intrinsic tissue characteristics that are robust to transient pathological changes. Building on this concept, we adapted the habitat analysis strategy to BSCMs to derive imaging biomarkers that remain consistent across different haemorrhage stages. By integrating these habitat-derived features with clinical variables, this study aims to: (1) construct an MRI habitat atlas that is robust across haemorrhage stages and standardises the representation of BSCM architecture and (2) develop and validate a multimodal predictive model for individualised assessment of re-haemorrhage risk.

Method

Participants and clinical data collection

The model was derived and internally validated using a retrospective cohort of consecutive adult patients with BSCM treated at Huashan Main Campus between January 2015 and December 2020. Eligible patients were randomly allocated in a 7:3 ratio to either a discovery set for model development or an internal validation set. External temporal validation was conducted using a prospective cohort from Huashan West Campus between January 2021 and June 2023.

Inclusion criteria comprised: MRI-confirmed BSCM, evidenced by a classic ‘popcorn’ appearance with a surrounding T2 hypointense hemosiderin rim³; a treatment-free observation period of at least 12 months after the baseline MRI or the occurrence of a subsequent haemorrhage within this period; and age ≥18 years at diagnosis. Key exclusion criteria included: symptomatic haemorrhage at presentation requiring immediate intervention; radiation-induced cavernous malformations; coagulopathy or ongoing anticoagulant/antiplatelet therapy (with the exception of low-dose aspirin); and Zabramski type IV lesions. The primary endpoint was symptomatic re-haemorrhage, defined as either radiologically confirmed haemorrhage.¹⁶

MRI acquisition and pre-processing

All scans were acquired on 3 T MRI systems (Siemens Verio and GE Signa Premier in Main Campus, Philips Ingenia CX in West Campus) using a standardised protocol comprising the following sequences: 3-D T1-weighted (TR 2.2–2.4 s, TE 2.9–3.1 ms, TI 850–900 ms, flip angle 8°–9°, 1 mm³ isotropic resolution), T2-weighted (TR 3.1–3.3 s, TE 250–260 ms, 1 mm³ isotropic), contrast-enhanced T1-weighted (identical parameters to pre-contrast T1, acquired 5 min after intravenous administration of 0.1 mmol kg⁻¹ gadolinium) and susceptibility-weighted imaging (SWI variants: TR 28–36 ms, TE 20–25 ms, flip angle 15°–20°, resolution 0.6×0.6×2 mm³). Raw DICOM images were converted to NIfTI format, resampled to an isotropic 0.5 mm resolution using bilinear interpolation, intensity-normalised to a range of 0–255, and rigidly co-registered across all sequences using ANTs. Two senior neurosurgeons, blinded to clinical information, performed manual 3D segmentation of the entire BCM using ITK-SNAP (V.4.0).

Spatial habitat segmentation and clustering

Figure 1 schematically depicts the computational pipeline for delineating intralesional spatial niches. After rigid registration and isotropic resampling (0.4 mm³) of each subject’s T2-weighted, contrast-enhanced T1-weighted and SWI, the T2 sequence, chosen for its superior sensitivity to brain-stem cavernous malformations, served as the reference for lesion masking. Within this regions of interest (ROI), every voxel’s multimodal feature vector was assembled into a high-dimensional matrix. The lesion was first over-segmented into a few hundred spatially coherent superpixels via K-means clustering, reducing computational burden while preserving local texture boundaries. These superpixels were then merged by an unsupervised Louvain graph-clustering algorithm implemented in Seurat 4.0, which, by optimising modularity, grouped units with similar multiparametric profiles into a minimal set of unified habitats, each representing a biologically consistent subregion with homogeneous haemorrhage risk.

Figure 1. Workflow for developing a combined imaging-clinical model to predict symptomatic re-haemorrhage in brainstem cavernous malformations. After excluding asymptomatic cases, co-registered multiparametric MRI (including T2-weighted, susceptibility-weighted imaging and contrast-enhanced T1-weighted sequences) and clinical data were processed through a pipeline comprising rigid registration, isotropic resampling and K-means-based habitat segmentation into biologically coherent subregions. Radiomic features were then extracted from these habitats, and the most informative imaging signatures were selected and integrated with clinically independent predictors to construct and internally validate a prognostic nomogram for re-haemorrhage risk.

To assess the temporal consistency of the spatial habitat distributions, the normalised cross-correlation (NCC) was used to quantify the similarity between multiregional spatial interaction (MSI) matrices derived from the same patient at different time points. The NCC between two matrices M¹ and M² is defined as:

$${\displaystyle {\displaystyle {\displaystyle \text{NCC}\left(M^1,M^2\right)=\frac{\sum _{i,j}{M}_{i,j}^1{M}_{i,j}^2}{\sqrt{\sum _{i,j}{M}_{i,j}^1{}^2}\sqrt{\sum _{i,j}{M}_{i,j}^2{}^2}}}}}$$

This metric ranges from 0 to 1, with values closer to 1 indicating higher similarity. For each patient, pairwise NCC values were computed between all available follow-up MSI matrices and then averaged. An average NCC value above 0.75 was considered indicative of good temporal consistency, confirming that the habitat-derived features remained stable across time despite the dynamic evolution of haemoglobin degradation products.

Radiomics feature extraction and preprocessing

Following the radiomics protocol established in our previous work,²¹ high-throughput quantitative features were extracted from T2-weighted images based on manually segmented lesions. Two experienced neurosurgeons performed the segmentation using ITK-SNAP to delineate each cavernous malformation. A comprehensive set of radiomics features was subsequently extracted, including shape descriptors along with texture and intensity features derived from first-order statistics and higher-order matrices (grey-level co-occurrence, run-length, size-zone, dependence and neighbouring grey-tone difference matrices). These features were computed from both the original images and a series of filtered representations (Gaussian, wavelet and edge-enhancement). To ensure reproducibility, only features with an intraclass correlation coefficient ≥0.75 were retained after intensity normalisation to the 0–1 range.

Model development and integration

Following feature extraction at both the habitat and conventional radiomic levels, we conducted two parallel modelling processes. Features derived from the MSI matrix were used to train seven candidate machine learning algorithms—SVM, LR, KNN, NB, DT, RF and XGBoost—employing threefold cross-validation and 1000 iterations of randomised hyperparameter optimisation. In parallel, high-dimensional conventional radiomic features underwent initial univariate screening via the Mann-Whitney U test, followed by Least Absolute Shrinkage and Selection Operator (LASSO) regression to identify the most discriminative subset and construct an optimised radiomics model. The resulting imaging risk score quantifies the predicted probability of BSCM haemorrhage. An imaging risk score was computed for each patient from the selected imaging model as the probability of having an unfavourable re-haemorrhage within 2 years and represented as a numeric value ranging from 0 to 1. This score was compared against conventional radiomics models to assess performance improvement. Subsequently, the Boruta algorithm was applied to identify the most relevant clinical predictors. These selected clinical features were integrated with the habitat-derived imaging risk score to train a hybrid multimodal prognostic model, which was ultimately presented as a clinical nomogram. Details regarding the model can be accessed online (https://github.com/guihaohuai/hab_code).

Statistical analysis

Statistical analyses were performed using R software (V.4.5.0). Continuous and categorical variables were summarised using appropriate descriptive statistics and compared with t-tests, Mann-Whitney U, χ² or Fisher’s exact tests. Model development involved feature selection via LASSO regression and training of multiple machine learning algorithms optimised through threefold cross-validation and hyperparameter tuning. The optimal model was selected based on the area under the curve (AUC) performance. The imaging risk score was dichotomised using Youden’s index, and a multimodal logistic model integrating imaging and clinical predictors—selected via the Boruta algorithm—was constructed. Model performance was evaluated using AUC, calibration curves, Hosmer-Lemeshow testing and decision curve analysis in both internal and temporal validation cohorts, with a two-sided p value <0.05 considered significant.

Results

Patient characteristics

From the retrospective cohort (Huashan Main Campus, 2015–2020), 205 eligible patients were included and randomly split 7:3 into a discovery set (n=147) and an internal validation set (n=58). The prospective temporal validation cohort (Huashan West Campus, 2021–2023) consisted of 94 eligible patients (figure 2). The entire cohort was 46.5% female with a median age of 44 years (IQR 33–54). Temporal validation patients were older (46±14 vs 41±12 years, p=0.006), had a higher rate of mesencephalic lesions (27% vs 7%, p<0.001) and a higher rate of acute haemorrhage (21.5% vs 31.9%). The 2-year symptomatic re-haemorrhage rates were 53%, 50% and 42% across the discovery, internal validation and temporal validation sets, respectively (p=0.21). Sex, lesion multiplicity, body mass index (BMI) and other lesion characteristics showed no significant intergroup differences (all p≥0.34), supporting cohort homogeneity for subsequent modelling (table 1).

Figure 2. Flowchart of participant screening and exclusion, illustrating the derivation of the final cohorts for model discovery (derivation cohort, n=147) and internal validation (n=58) from the Huashan Main Campus cerebral cavernous malformation retrospective cohort (2015–2020), and for prospective temporal validation (n=94) from the Huashan West Campus BSCM cohort (2021–2023). AUC, area under the curve; BSCM, brainstem cavernous malformation; mRS, modified Rankin Scale; MSE, mean squared error; MSI, multiregional spatial interaction.

Table 1. Patient demographics and baseline characteristics.

Characteristic	Cohort			P value
	Training set N=147	Internal validation N=58	Temporal external validation N=94
Age, years	41±12	44±13	46±14	0.006*
Female sex, n (%)	66 (44.9%)	30 (51.7%)	43 (45.7%)	0.667†
BMI, kg/m2	23.38±5.16	23.30±2.60	23.25±2.94	0.968*
Hypertension, n (%)	25 (17.0%)	8 (13.8%)	19 (20.2%)	0.589†
Diabetes, n (%)	3 (2.0%)	1 (1.7%)	2 (2.1%)	>0.999‡
Size (T2), mean±SD (cm)	1.91±0.79	2.03±0.80	2.14±0.75	0.078*
Axial midpoint, n (%)	85 (57.8%)	35 (60.3%)	44 (46.8%)	0.158†
Multiple lesions, n (%)	16 (10.9%)	7 (12.1%)	9 (9.6%)	0.885†
DVA, n (%)	34 (23.1%)	9 (15.5%)	16 (17.0%)	0.340†
Lesion location, n (%)				<0.001†
Mesencephalic	10 (6.8%)	2 (3.4%)	25 (26.6%)
Pontomesencephalic	37 (25.2%)	17 (29.3%)	21 (22.3%)
Pontine	56 (38.1%)	19 (32.8%)	25 (26.6%)
Pontomedullary	32 (21.8%)	15 (25.9%)	13 (13.8%)
Medullary	12 (8.2%)	5 (8.6%)	10 (10.6%)
Zabramski grade, n (%)				0.542†
I	44 (29.9%)	21 (36.2%)	29 (30.9%)
II	66 (44.9%)	19 (32.8%)	43 (45.7%)
III	37 (25.2%)	18 (31.0%)	22 (23.4%)
Haemorrhage, n (%)				0.016†
Acute	25 (17.0%)	19 (32.8%)	30 (31.9%)
Subacute	86 (58.5%)	22 (37.9%)	46 (48.9%)
Chronic	36 (24.5%)	17 (29.3%)	18 (19.1%)
Re-haemorrhage, n (%)	78 (53.1%)	29 (50.0%)	39 (41.5%)	0.211†

DVA, developmental venous anomaly.

^*One-way analysis of variance.

^†Pearson’s χ2 test.

^‡Fisher’s exact test.

BMI, body mass index.

Habitat-driven phenotyping of BCMs via multiparametric MRI

As illustrated in figure 3A,B, each lesion was initially over-segmented into 200–400 spatially contiguous superpixels using K-means clustering. These superpixels were subsequently aggregated into six biologically distinct habitats through unsupervised clustering with the Louvain algorithm implemented in Seurat 4.0. The optimal number of clusters was determined using the Elbow Method, which evaluates clustering performance by computing the Within-Cluster Sum of Squares (WCSS) for varying values of k. The WCSS is defined as:

Figure 3. MRI-based habitat clustering and its association with re-bleeding risk in brainstem cavernous malformations. (A, B) UMAP visualisation of voxel-wise clustering derived from multiparametric MRI, illustrating distinct intralesional subregions. (C) Feature characteristics of the six identified clusters, including T2 zone percentage, run-length non-uniformity, entropy and T1 intensity properties. (D) Six representative cases showing habitat distribution: the top three rows depict patients without re-bleeding, where clusters 1 and 2 are predominant; the bottom three rows represent patients who experienced re-bleeding within a 2-year follow-up, with clusters 3–6 (particularly clusters 4 and 5) comprising the major habitat components. These findings suggest that specific MRI-derived habitats may be associated with a higher risk of haemorrhage in cavernous malformations. UMAP, uniform manifold approximation and projection.

$${\displaystyle {\displaystyle {\displaystyle \text{WCSS}\left(k\right)=\sum _{i=1}^k\sum _{x\in C_i}||x-{\mu }_i|{|}^2}}}$$

where k denotes the number of clusters, $C_i$ represents the ith cluster, ${\mu }_i$ is the centroid of cluster i and x refers to a data point within the cluster. We observed that beyond k=6, the rate of reduction in WCSS markedly decreased, indicating diminishing returns in cluster compactness with additional subdivisions. Thus, k=6 was selected as the optimal number, ensuring both biological interpretability and computational efficiency.

Guided by the MSI framework, we extracted 128 quantitative descriptors from these habitats (online supplemental table S1), encompassing absolute and relative subregion volumes, boundary interactions and second-order texture features derived from Gray Level Co-occurrence Matrix (GLCM), Gray-Level Run-Length Matrix (GLRLM), Gray-Level Size Zone Matrix (GLSZM) and Gray-Level Dependence Matrix (GLDM) analyses. Uniform manifold approximation and projection (UMAP) dimensionality reduction projected the lesions into six non-overlapping clusters, each exhibiting highly reproducible radiological signatures (figure 3C,D). For instance, cluster 1 was characterised by significantly elevated T2_ZonePercentage alongside suppressed T1_Maximum and T1_90Percentile, indicative of chronic hemosiderin deposition. In contrast, cluster 6 showed prominent increases in T1_90Percentile and T1_DependenceEntropy, consistent with acute/subacute haemorrhage.

Model selection and performance

Among the seven machine learning models evaluated, XGBoost demonstrated superior performance, achieving an AUC of 0.81 (95% CI 0.78 to 0.85) in the discovery set, with consistent discrimination in the internal validation (AUC=0.80) and temporal external validation cohorts (AUC=0.82) (figure 4A, table 2). Shapley additive explanations (SHAP) analysis identified MSI116_gldm_DependenceNonUniformity and MSI69_glcm_ldmn as among the most influential predictive features (figure 4B). Notably, the feature importance profile was highly consistent in the external validation set (figure 4C), with the XGBoost-based imaging risk score showing clear separation between haemorrhage and non-haemorrhage groups across all cohorts. Furthermore, the model significantly outperformed a conventional radiomics approach (AUC 0.82 vs 0.73; online supplemental figure S1), underscoring its robustness and generalisability for BSCM risk assessment.

Figure 4. Performance and interpretability of the XGBoost model for re-haemorrhage prediction using multiregional spatial interaction (MSI) features. The model demonstrated strong discrimination across cohorts, with AUCs of 0.81 (95% CI 0.78 to 0.85) in the discovery set, 0.80 in the internal validation set and 0.82 in the temporal external validation set, as shown by ROC curves (A). SHAP analysis (B) revealed the top predictive MSI features and their directional impact on model output. Individual re-haemorrhage risk scores derived from the model (C) were significantly higher in patients with re-haemorrhage across all cohorts, illustrating the model’s consistent stratification ability. (D) Generalised additive model smooth curve showing the linear relationship between the imaging risk score and the log(odds) of re-haemorrhage, after adjusting for clinical covariates. AUC, area under the curve; ROC, receiver operating characteristic; SHAP, shapley additive explanations.

Table 2. Evaluation of candidate machine learning imaging models via cross-validation in the training set.

Model	AUC	Accuracy	Sensitivity	Specificity	F1
SVM	0.73	0.76	0.41	0.92	0.52
LR	0.68	0.77	0.41	0.93	0.53
KNN	0.59	0.64	0.50	0.70	0.47
NB	0.73	0.76	0.41	0.92	0.52
DT	0.77	0.68	0.56	0.74	0.53
RF	0.78	0.76	0.65	0.81	0.63
XGBoost	0.81	0.78	0.62	0.80	0.61

AUC, area under the curve.

A generalised additive model, adjusted for gender, age, BMI and modified Rankin Scale (mRS), confirmed a strictly linear relationship between the imaging risk score and log-odds of re-haemorrhage (edf≈1.0, p<0.001; figure 4D), with no evidence of threshold or saturation effects. This linear dose–response association supports the reliability of the score as a continuous predictor of re-haemorrhage risk.

Derivation of imaging risk groups

To facilitate clinical translation, we established a simplified imaging-based grading system (figure 5A). In the derivation cohort, high-risk lesions exhibited a 2-year cumulative re-haemorrhage rate of 81% (101/125), compared with 26% (45/174) in low-risk lesions (log-rank p<0.001). This binary imaging classifier achieved an AUC of 0.77 (95% CI 0.72 to 0.82), with a positive predictive value of 0.81 and a negative predictive value of 0.74 (online supplemental figure S2A). Calibration curves showed excellent agreement between predicted and observed outcomes (Hosmer-Lemeshow p=0.29), and decision-curve analysis demonstrated sustained net clinical benefit across risk thresholds ranging from 5% to 85%.

Figure 5. Development, interpretation and validation of a multimodal clinical-imaging nomogram for re-haemorrhage risk prediction. The optimal risk score cut-off (0.535) was determined using Youden’s index to stratify patients into high- and low-risk re-haemorrhage groups (A). Key clinical predictors were selected via the Boruta algorithm for inclusion in the combined model (B). A comprehensive nomogram was constructed incorporating these selected clinical and imaging-derived predictors, with bar length and density reflecting each variable’s contribution and distribution (C). Receiver operating characteristic analysis demonstrated significantly superior performance of the combined clinical-imaging model (AUC=0.851 in temporal external validation; 0.833 in training) compared with the clinical-only model (AUC=0.657–0.678), with Delong tests confirming statistically significant improvement (p<0.001) (D). AUC, area under the curve; BMI, body mass index; mRS, modified Rankin Scale.

Evaluation of multimodal prognostic models

The imaging risk grade was integrated with three clinical predictors—female sex (OR 2.82, 95% CI 1.42 to 5.80), BMI (OR 1.10, 95% CI 0.99 to 1.25) and admission mRS score (OR 2.11, 95% CI 1.34 to 3.41)—to develop a multimodal logistic model (figure 5B–D). The combined model achieved an AUC of 0.833 (95% CI 0.777 to 0.889) in the training set, significantly outperforming both the imaging-only (ΔAUC+0.07, p=0.008) and clinical-only models (ΔAUC+0.17, p<0.001) (figure 5E). Internal validation via bootstrapping confirmed strong discrimination (optimism-corrected C-index: 0.816). Decision-curve analysis indicated superior net benefit within the clinically relevant threshold range of 5%–75% (online supplemental figure S2B,C).

Subgroup analysis based on multimodal risk stratification revealed divergent predictors of re-haemorrhage. In the low-risk subgroup (n=174), female sex (adjusted OR 2.89, 95% CI 1.32 to 6.34; p=0.008), larger lesion size (adjusted OR 1.87, 95% CI 1.08 to 3.24; p=0.025) and higher mRS scores (adjusted OR 1.84, 95% CI 1.20 to 2.82; p=0.005) were significantly associated with re-haemorrhage (online supplemental figure S3A). In contrast, none of the clinical variables showed significant association in the high-risk subgroup (n=125; all p>0.05; online supplemental figure S3B).

Discussion

In this prospective cohort of 299 adults with BSCM, the 2-year cumulative re-haemorrhage rates following an initial bleed were 26% in the low-risk stratum and 81% in the high-risk stratum. The imaging-based stratification alone achieved an AUC of 0.77, which improved to 0.84 after incorporating female sex, BMI and baseline mRS scores. This enhancement underscores the inadequacy of relying solely on binary prior-haemorrhage labels and highlights the necessity of multimodal risk assessment. Given that both microsurgery and radiosurgery entail a 5%–35% risk of permanent neurological deficits,^22–24 the ability to prospectively identify patients with a high likelihood of recurrent haemorrhage yet low operative morbidity can significantly refine the timing and approach of intervention.

Conventional radiomics models, including our prior efforts, have shown utility in predicting surgical outcomes but are limited by the dynamic evolution of haemoglobin degradation products, which introduces substantial temporal variability in texture features. Moreover, whole-lesion ROIs often conflate haematoma with the vascular nidus, and machine learning models are prone to overfitting, particularly with moderate sample sizes and across different MRI sequences. To address these challenges, we introduced, for the first time, a tumour-habitat atlas approach to BSCM. Using unsupervised clustering, each lesion was partitioned into six biologically coherent subregions. Radiomic features extracted from these habitats—minimally influenced by haemorrhagic artefacts—exhibited over 90% concordance in UMAP visualisations across follow-up scans for both high- and low-risk patients, providing a stable imaging signature despite temporal changes in intralesional haemorrhage constituents (figure 6).

Figure 6. Illustrative cases demonstrating the application of the MRI habitat atlas and dynamic nomogram for individualised re-haemorrhage risk prediction over time. (A) A 59-year-old male presenting with limb numbness and an initial mRS of 4 experienced symptomatic re-haemorrhages at 3 and 5 months, confirmed by MRI. Habitat analysis across three time points showed consistent enrichment of clusters 4–6 (mean normalised cross-correlation, NCC=0.85), indicating high-risk vascular instability. The dynamic nomogram predicted a 2-year re-haemorrhage risk of 0.895. (B) A 48-year-old male with diplopia and limb weakness (initial mRS=4) showed haemorrhage resolution on follow-up MRI at 1 and 2 years, with no recurrent symptomatic bleeding. Habitat mapping revealed predominant clusters 1–3 (mainly 2–3; mean NCC=0.82), suggesting stable evolution. The nomogram predicted a 2-year re-haemorrhage risk of 0.456. BMI, body mass index; mRS, modified Rankin Scale.

We developed a two-step individualised prediction framework centred on a habitat-derived risk score, further integrated with BMI, gender and admission mRS within a logistic model for re-haemorrhage risk. Although we did not develop an ordinal model to assess the degree of neurological deterioration following subsequent haemorrhage, the current model effectively stratifies re-haemorrhage risk and has been validated both internally and externally. Although subgroup analyses reaffirmed the established roles of female sex, larger lesion size and higher baseline mRS as clinical risk factors for re-haemorrhage,^{6 13 25} a model relying exclusively on these variables demonstrated limited and inconsistent predictive performance across cohorts. This emphasises the critical constraint of using clinical metrics alone, which may be confounded by population heterogeneity and temporal dynamics in haemorrhage progression. Our multimodal habitat atlas effectively mitigates these issues by capturing intrinsic imaging biomarkers of haemorrhage risk, thereby offering a more robust and reproducible predictive tool.

Based on this model, patients can be stratified into three management categories: those with high haemorrhage and high deterioration risk may be candidates for early surgery or stereotactic radiosurgery; those with high haemorrhage risk but low deterioration risk may benefit from individualised risk–benefit assessment regarding observation versus intervention; and those with low haemorrhage risk may be managed conservatively with imaging surveillance and symptomatic monitoring to avoid overtreatment.

Several limitations should be acknowledged. This study is a single-centre observational investigation, with a partial external validation cohort conducted prospectively. Although the model demonstrates good generalisability, the inherent referral bias of a quaternary neurosurgical centre may lead to overestimation of natural-history events. MRI time points and sequences were not fully standardised across all patients. The number of habitats was fixed a priori at six, and external validation against histopathological or genomic correlates was not performed. Although the re-haemorrhage nomogram underwent 10-fold cross-validation, it has yet to be tested in a fully independent, prospective, multicentre cohort. Calibration curves also indicated mild overfitting at extreme risk values. Additionally, medication exposure—specifically the use of antiplatelet or anticoagulant agents such as aspirin and clopidogrel—was not quantitatively correlated with the timing of haemorrhage events, highlighting the need for prospective studies with controlled designs to address this gap.

Conclusion

We developed an MRI habitat atlas for BSCMs that is robust across haemorrhage stages and integrated habitat-derived biomarkers with clinical variables into a multimodal risk model. This tool accurately predicts 2-year re-haemorrhage risk, consistently identifies high-risk lesions independent of haemoglobin degradation product evolution and supports personalised clinical decision-making to avoid overtreatment.

Data availability statement

Data are available upon reasonable request.