Predicting local control of brain metastases after Gamma Knife radiosurgery: a multimodal deep learning-radiomics approach

Abstract

Background

Accurate prediction of local control (LC) following Gamma Knife radiosurgery (GKRS) remains clinically challenging for patients with brain metastases (BMs) secondary to breast or lung cancer. The scarcity of validated pre-therapeutic biomarkers currently hinders the development of individualized treatment decision-making strategies.

Methods

This retrospective analysis included 324 BMs from 195 patients. To develop and benchmark various predictive frameworks, several models were constructed: clinical-only, radiomics-only, 2D deep learning (DL)-only, fused radiomics-DL (DLR), and a Combined model integrating all three modalities (clinical, radiomic, and DL features). All imaging signatures were derived from pre-treatment contrast-enhanced T1-weighted (T1CE) magnetic resonance imaging (MRI). Model performance was evaluated using the area under the curve (AUC), calibration plots, the Hosmer-Lemeshow test, and decision curve analysis (DCA).

Results

The Combined model demonstrated consistently high numerical discriminative performance, achieving an AUC of 0.854 (95% Confidence Interval [CI]: 0.791–0.917) in the training cohort and 0.781 (95% CI: 0.618–0.945) in the independent test cohort. DCA demonstrated that the Combined model yielded superior net clinical benefit across threshold probabilities ranging from 10% to 50%, outperforming all other candidate models. Consequently, a nomogram based on the Combined model was developed to facilitate individualized risk estimation.

Conclusions

The integration of clinical parameters, handcrafted radiomic features, and 2D deep learning signatures provides a robust predictive tool for LC following GKRS in patients with brain metastases from breast or lung cancer. This multimodal approach holds significant potential for refining post-radiosurgical management and advancing personalized neuro-oncology care.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02579-7.

Background

Brain metastases (BMs) represent the most prevalent intracranial neoplasms in adults, with their incidence continuously escalating due to advancements in systemic treatments and neuroimaging methodologies [1]. The most common primary sites for metastatic brain disease are breast cancer and non-small cell lung cancer (NSCLC). Specifically, 20–40% of NSCLC patients and 10–30% of breast cancer patients, particularly those with HER2-amplified or triple-negative subtypes, develop BMs during their disease course [2–4]. The development of BMs leads to substantial neurological morbidity and remains a leading cause of mortality, underscoring the critical need for effective local therapeutic interventions [5].

The therapeutic paradigm for limited BMs has shifted from whole-brain radiotherapy (WBRT), which is now largely obsolete due to its neurocognitive side effects, to more precise focal irradiation techniques such as stereotactic radiosurgery (SRS) or Gamma Knife radiosurgery (GKRS) [6, 7]. For patients presenting with 5–20 cerebral metastases, current evidence favors SRS over hippocampal-sparing WBRT, as it effectively alleviates symptoms and preserves cognitive function and quality of life [8]. Nevertheless, clinical outcomes remain highly heterogeneous; while some lesions achieve durable control, others experience local failure (LF) within 12 months [9, 10]. Currently, the lack of reliable pre-treatment biomarkers to predict these divergent responses represents a significant challenge in personalized neuro-oncology.

Standard response assessment relies on serial contrast-enhanced T1-weighted (T1CE) magnetic resonance imaging (MRI) and size-based criteria, notably the Response Assessment in Neuro-Oncology for Brain Metastases (RANO-BM) guidelines [11]. However, this approach is inherently reactive and often confounded by radiation-induced pseudoprogression, which can mimic true disease progression and delay essential salvage interventions [12, 13]. Given that modern systemic therapies have extended median survival beyond previous horizons, a pre-treatment biomarker capable of predicting LC shortly after GKRS would facilitate risk-stratified follow-up, earlier treatment intensification, and enhanced resource allocation [1].

Radiomics, an emerging field characterized by the high-throughput extraction of quantitative imaging features, has enabled the identification of distinct tumor phenotypes associated with therapeutic response across various malignancies [14, 15]. Concurrently, deep learning (DL) techniques, which derive hierarchical representations from raw image data, frequently surpass traditional feature sets in predictive tasks [16, 17]. Nevertheless, utilizing these modalities in isolation provides only a fragmented understanding of tumor biology. Preliminary research indicates that the integration of radiomic and DL features may enhance prediction accuracy. However, this approach has not been thoroughly investigated in the context of predicting GKRS outcomes for breast and lung cancer brain metastases. Integrating these imaging biomarkers with clinical factors, such as metastasis volume, primary histology, and performance status, is essential to enable actionable decision support [17, 18].

Therefore, this study aimed to develop a multimodal framework for predicting local control following GKRS in patients with brain metastases originating from breast or lung primaries. While existing literature has explored various predictive models, our work distinguishes itself by providing a comprehensive head-to-head evaluation of five distinct predictive model architectures (clinical-only, radiomics-only, DL-only, fused radiomics-DL, and a fully Combined model) within a single, homogeneous dataset. Radiomic and deep learning features were extracted from pre-treatment T1CE MRI, then refined using correlation filtering and least absolute shrinkage and selection operator (LASSO) regression. The Combined model consistently demonstrated superior discriminative performance in an independent test cohort, offering the highest clinical utility for individualized risk estimation and guiding post-radiosurgical management. Crucially, our study also rigorously assessed the clinical utility of these models using Decision Curve Analysis (DCA) and translated the optimal Combined model into an interpretable nomogram, aiming to provide a more transparent and potentially actionable tool for personalized neuro-oncology care. Ultimately, the findings of this study underscore that the integration of handcrafted radiomics with 2D deep learning features significantly enhances the prediction of GKRS response across two primary histologies of brain metastases, thereby making a substantial contribution to the existing body of evidence in this field.

Methods

Study population and case selection

This retrospective study enrolled patients who underwent GKRS for metastatic brain lesions at our institution between November 2019 and March 2025 (Fig. 1). Inclusion criteria were: (a) pathologically confirmed primary malignancy of breast or lung origin, verified by biopsy or surgical resection; (b) complete and accessible medical records with sufficient follow-up data; (c) MRI-confirmed intracranial metastases; and (d) GKRS administered at our center. Exclusion criteria included: the presence of more than five cerebral lesions or the absence of a six-month post-GKRS MRI evaluation. The rationale for excluding patients with more than five cerebral lesions was based on the prevailing clinical consensus for oligometastatic disease during the study design period (2019) and aimed to enhance cohort homogeneity for robust predictive model development, aligning with similar prior modeling studies [19]. After applying these selection criteria, 195 patients contributing 324 individual brain metastases were included in the analysis. Pretreatment contrast‑enhanced cerebral MR images were acquired using a 3.0T Philips Ingenia scanner. Serial post‑treatment imaging was performed at 3, 6, 9, and 12 months post-GKRS. Disease progression, defined as LF, was characterized by a relative increase in tumor volume on the six‑month follow‑up MRI compared to the baseline scan. Local control (LC) was defined as one of three outcomes: no enlargement, volumetric reduction, or complete disappearance of the treated lesion. These response definitions align with the RANO‑BM guidelines [20]. Given the potential for each metastatic focus to exhibit an independent treatment response, the entire lesion set was randomly partitioned into a training subset (70% of lesions, n = 228) and a separate test subset (30%, n = 96). Figure 2 provides a schematic overview of the study workflow. The Ethics Review Board of the First Affiliated Hospital of Nanjing Medical University granted approval for this investigation (Approval No. 2025‑SR‑1239).

Fig. 1.

Flowchart of patient selection and exclusion criteria. GKRS, Gamma Knife radiosurgery

Fig. 2.

Multimodal workflow for local control prediction. This schematic illustrates the comprehensive framework for predicting local control (LC) after Gamma Knife radiosurgery (GKRS) in brain metastases. The workflow integrates three parallel feature streams: (i) Radiomics (Rad) Feature Generation: Features are extracted from pre-treatment T1-CE MRI and processed to form the Radiomics signature. (ii) Deep Learning (DL) Feature Extraction: Deep features are derived from the largest tumor axial slice of T1-CE MRI to create the Deep Learning signature. (iii) Clinical Feature Selection: Significant clinical variables are selected to establish the Clinical signature. These signatures are then integrated: Rad and DL features are fused to form the Deep Learning Radiomics (DLR) model, which is subsequently combined with the Clinical signature to construct the final Combined model. All models are evaluated on an independent test cohort

Clinical parameters

A comprehensive set of clinical and therapeutic variables was collected. These included patient sex, primary tumor histology, metastatic volume, lesion location, prescribed radiation dose, age at primary diagnosis and at brain metastasis diagnosis, prior treatments (cranial surgery, prior SRS, WBRT, systemic therapy), status of extracranial or nodal metastases, and Karnofsky Performance Status (KPS). Tumor volumes were calculated based on initial manual segmentations. The prescription dose was defined as the radiation dose administered to the target lesions margin.

MR image acquisition and target delineation

Structural MR images were acquired with patients positioned in a stereotactic head frame using a 3.0T scanner. This study exclusively utilized axial T1CE sequences, which were retrieved as DICOM files from our institutional Picture Archiving and Communication System (PACS; Carestream, Canada). For the training cohort, each metastasis region of interest (ROI) was manually delineated on T1 CE sequences using ITK-SNAP software [21]. Two experienced radiologists independently performed the delineations, and any discrepancies were resolved through consultation with a senior neuroradiologist possessing over 20 years of experience. To assess segmentation consistency, all ROIs were re-outlined by the same readers after a two-month interval. These segmentations were subsequently used to compute intraclass correlation coefficients (ICCs) for both inter-observer and intra-observer reliability.

Radiomics feature generation

Radiomic features were subsequently computed and categorized into three categories: geometric (morphological), intensity-based (first-order), and textural (higher-order). Geometric features are known to effectively quantify the three-dimensional shape characteristics of each tumor. Intensity features, reflecting first-order voxel statistics, provide insights into intra-tumoral homogeneity or heterogeneity. Textural features, conversely, describe higher-order spatial patterns of intensity distribution, and were derived using techniques such as the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size-zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM). In total, 1,197 features were extracted, comprising 14 geometric, 234 intensity, and 949 textural features. A comprehensive list of the extracted features is detailed in Supplementary Material 1. All feature calculations were performed using an in-house pipeline developed based on the PyRadiomics platform (http://pyradiomics.readthedocs.io).

Feature dimensionality reduction and classifier training

Normalization of all radiomic features was conducted through z-score standardization. To address high inter-feature collinearity (Pearson correlation coefficient ≥ 0.9), a greedy recursive elimination strategy was employed. This involved iteratively removing features whose correlations exceeded the predefined threshold until no strongly correlated pairs persisted. Subsequently, features exhibiting high reproducibility, defined by ICC > 0.75 for both inter-observer and intra-observer analyses, were selected. Following this, LASSO regression analysis was applied to further reduce the number of features, resulting in 8 radiomic features retained for subsequent analysis. For deep learning features, Principal Component Analysis (PCA) was implemented to retain the top 8 principal components from the 1024-dimensional feature vectors, providing a comprehensive summary of the most informative deep learning features per sample. A variety of machine learning classifiers were then trained to develop radiomics-based risk models, including logistic regression (LR), support vector machine (SVM), random forest, extra trees, and light gradient boosting machine (LightGBM). For each classifier, systematic hyperparameter optimization was conducted using GridSearchCV with 5-fold cross-validation on the training cohort. Optimal hyperparameters were determined by maximizing the Area Under the Curve (AUC) score.

The predictive performance of these models was systematically compared. Furthermore, the incremental value of integrating multimodal imaging signatures was assessed to determine their contribution to enhanced predictive accuracy (see Fig. 3).

Fig. 3.

Radiomics (Rad) signature development and performance. (A) LASSO coefficient paths from five-fold cross-validation in the training cohort. (B) Bar plot showing the distribution of Rad scores based on the selected features. (C, D) Comparison of predicted probabilities from the Rad model in the training (C) and test (D) cohorts

Two‑dimensional deep learning framework and feature extraction

For the 2D deep learning model, the axial slice exhibiting the largest ROI cross-sectional area was selected as the representative image. To minimize background interference, each ROI was manually cropped to its minimal bounding rectangle. Before network processing, all images underwent channel-wise z-score normalization to standardize intensity distributions. During the training phase, online data augmentation techniques, including random cropping, horizontal flipping, and vertical flipping, were applied. For inference, test images were processed solely with normalization.

This study employed a deep learning architecture based on DenseNet121 [22, 23]. To improve transferability across diverse patient demographics, transfer learning was utilized, initializing the network with weights pre-trained on the ImageNet dataset. A cosine annealing learning rate schedule was adopted to promote model convergence and enhance generalization, as expressed by the following formula:

1

In this context, the minimum and maximum learning rates are denoted by and , respectively. The total number of training epochs is denoted by =50. The model underwent optimization through the application of stochastic gradient descent (SGD) with a softmax cross-entropy loss function, an initial learning rate of 0.01, a batch size of 32, Nesterov momentum of 0.9, and a weight decay of . Early stopping with a patience of 10 epochs was also employed.

A fused feature representation was created by concatenating the deep learning features () with the radiomics features ():

2

From the final convolutional layer of the pre-trained DenseNet121 model, 1,024-dimensional feature vectors were extracted, thus deriving features within a deep learning framework. To reduce dimensionality and address potential multicollinearity, the PCA method was implemented. This aimed to retain eight principal components that comprehensively summarized the most informative deep learning features per sample. The output from the final softmax layer of the convolutional neural network (CNN) was subsequently designated as the deep learning probability signature for further integration.

Deep Learning Radiomics (DLR) signature construction

The DLR signature was constructed by concatenating the extracted deep learning features and the handcrafted radiomic features at the input level (pre-fusion). The combined feature set then underwent the same feature selection and model training pipeline previously described for the radiomics-only model.

Clinical‑only model development

The clinical model was developed using clinical and conventional radiological variables, as enumerated in the Supplementary Material 2. All features were standardized via z-score normalization. Statistically significant differences between the LC and LF groups in the training set were identified (p < 0.05, assessed by independent samples t-test or Mann-Whitney U test, as appropriate) and subsequently retained for analysis. These selected features were then used to train a machine learning classifier, following the procedure outlined in the “Feature Dimensionality Reduction and Classifier Training” section (see Fig. S1). The specific clinical-radiological variables chosen for the Clinical Signature are further detailed in the “Results” section.

Integrated multimodal combined model

The Combined model was developed by integrating the selected significant clinical features with the output probability (deep learning probability score) from the DLR model using logistic regression. This composite predictor, termed the Combined Signature, was then represented as a nomogram for individualized risk prediction. The model’s performance underwent rigorous evaluation using a separate test cohort. Discriminative ability was assessed through receiver operating characteristic (ROC) curves and quantified by the area under the curve (AUC). Model calibration was evaluated using calibration plots and the Hosmer-Lemeshow (HL) goodness-of-fit test. Furthermore, the clinical utility was quantified via DCA to estimate net benefit across a range of clinically relevant probability thresholds.

Statistical procedures

For continuous variables, comparisons were performed using either the t-test or the Mann-Whitney U test, following a normality assessment via the Shapiro-Wilk test. Categorical variables were analyzed using Chi-square (χ²) tests. All statistical computations were executed using Python 3.7.12, with statistical evaluations conducted using Statsmodels version 0.13.2. Radiomic feature extraction was performed using PyRadiomics version 3.0.1. Machine learning algorithms were implemented using Scikit-learn version 1.0.2. The development of these frameworks was facilitated by PyTorch version 1.11.0, optimized for performance with CUDA version 11.3.1 and cuDNN version 8.2.1.

Results

Patient characteristics

A total of 195 patients with 324 brain metastases were enrolled in this retrospective analysis. Among these, 121 patients (62.1%) presented with a single metastatic lesion, while 74 patients (37.9%) had multiple lesions (ranging from 2 to 5 per patient). Specifically, 40, 18, 11, and 5 patients contributed 2, 3, 4, and 5 lesions, respectively. The entire lesion dataset was methodically partitioned into a training set (228 lesions, 70%) and a test set (96 lesions, 30%). As detailed in Table 1, no statistically significant differences in baseline demographic or clinical characteristics were observed between these two cohorts (all p values > 0.05), confirming that the random division resulted in well-balanced subsets. The median tumor volume was 200 mm³ (interquartile range [IQR]: 75.25–613.5 mm³). Supratentorial localization was the predominant lesion type, accounting for 69.14% of the observed lesions. At the pre-specified six-month MRI follow-up, 283 metastases (87.35%) met the criteria for LC, while the remaining 41 lesions (12.65%) exhibited LF. A comparative analysis within the training cohort revealed that four clinical-radiological variables—namely, prescribed treatment dose, prior brain treatment, prior WBRT, and presence of extracranial metastases—differed significantly between the LC and LF groups (Supplementary Material 2). The subsequent selection of these four parameters was made for the purpose of constructing the Clinical Signature.

Table 1.

Baseline demographic and clinical characteristics of the study population, stratified by training and test cohorts

Characteristics	All (n = 324)	Training set (n = 228)	Test set (n = 96)	p value
Volume of metastases# (mm3)	200(75.25, 613.5)	194(72.5, 662.75)	213(80.25, 386.25)	0.960
Treatment dose# (Gy)	15.60 ± 1.85	15.68 ± 1.91	15.42 ± 1.68	0.311
Age at diagnosis of brain metastases# (years)	56.73 ± 9.91	57.43 ± 9.84	55.04 ± 9.91	0.071
Age at diagnosis of primary cancer# (years)	53.61 ± 10.80	54.30 ± 10.74	51.97 ± 10.83	0.087
Sex*				0.667
Male	88(27.16)	64(28.07)	24(25.00)
Female	236(72.84)	164(71.93)	72(75.00)
Histology*				0.235
Lung cancer	170(52.47)	125(54.82)	45(46.88)
Breast cancer	154(47.53)	103(45.18)	51(53.12)
Location*				0.308
Infratentorium	100(30.86)	66(28.95)	34(35.42)
Supratentorium	224(69.14)	162(71.05)	62(64.58)
Prior brain treatment*				0.655
No	260(80.25)	181(79.39)	79(82.29)
Yes	64(19.75)	47(20.61)	17(17.71)
Prior SRS*				0.338
No	302(93.21)	215(94.30)	87(90.62)
Yes	22(6.79)	13(5.70)	9(9.38)
Prior WBRT*				0.576
No	287(88.58)	200(87.72)	87(90.62)
Yes	37(11.42)	28(12.28)	9(9.38)
Prior surgery*				1.000
No	313(96.60)	220(96.49)	93(96.88)
Yes	11(3.40)	8(3.51)	3(3.12)
Prior systemic treatment*				0.307
No	13(4.01)	7(3.07)	6(6.25)
Yes	311(95.99)	221(96.93)	90(93.75)
Presence of extracranial metastases*				0.725
No	56(17.28)	41(17.98)	15(15.62)
Yes	268(82.72)	187(82.02)	81(84.38)
Presence of lymph node metastases*				1.000
No	76(23.46)	53(23.25)	23(23.96)
Yes	248(76.54)	175(76.75)	73(76.04)
KPS score*				0.754
40	10(3.09)	9(3.95)	1(1.04)
50	8(2.47)	6(2.63)	2(2.08)
60	70(21.60)	48(21.05)	22(22.92)
70	176(54.32)	122(53.51)	54(56.25)
80	58(17.90)	42(18.42)	16(16.67)
90	2(0.62)	1(0.44)	1(1.04)

^#Continuous variables are presented as mean ± standard deviation or median with interquartile range, and were compared using the independent‑samples t‑test or Mann‑Whitney U test as appropriate. ^*Categorical variables are expressed as counts (n) with percentages, and between‑group differences were assessed using the chi‑square test or Fisher’s exact test. Abbreviations: SRS, stereotactic radiosurgery; WBRT, whole‑brain radiotherapy; KPS, Karnofsky Performance Status

Performance of the Deep Learning Radiomics (DLR) signature

Table 2 presents the predictive performance of the DLR signature, evaluated using five distinct machine learning classifiers. In the training cohort, the RandomForest classifier demonstrated the highest AUC of 0.972 (95% Confidence Interval [CI]: 0.946–0.997). It was followed by ExtraTrees (AUC = 0.915, 95% CI: 0.857–0.972), LR (AUC = 0.783, 95% CI: 0.695–0.870), LightGBM (AUC = 0.738, 95% CI: 0.638–0.838), and SVM (AUC = 0.668, 95% CI: 0.586–0.751). For the independent test cohort (Fig. 4), the corresponding AUC values were: LR, 0.774 (95% CI: 0.630–0.918); ExtraTrees, 0.749 (95% CI: 0.612–0.887); RandomForest, 0.726 (95% CI: 0.593–0.860); LightGBM, 0.695 (95% CI: 0.531–0.859); and SVM, 0.634 (95% CI: 0.508–0.760). A consistent decline in performance from training to test sets was observed across all classifiers, particularly for the RandomForest and ExtraTrees models, indicating some degree of overfitting. For the DLR signature, the AUC decreased from 0.783 in the training cohort to 0.774 in the independent test cohort when using logistic regression, a trend further elaborated in the Discussion section.

Table 2.

Comparison of model performance based on Deep Learning Radiomics Signature using different machine learning algorithms

Model	Accuracy	AUC	95% CI	Sensitivity	Specificity	PPV	NPV	Cohort
LR	0.662	0.783	0.695–0.870	0.815	0.642	0.234	0.963	Training
LR	0.823	0.774	0.630–0.918	0.643	0.854	0.429	0.933	Test
SVM	0.557	0.668	0.586–0.751	0.815	0.522	0.186	0.955	Training
SVM	0.531	0.634	0.508–0.760	0.786	0.488	0.208	0.930	Test
RandomForest	0.930	0.972	0.946–0.997	0.963	0.925	0.634	0.995	Training
RandomForest	0.729	0.726	0.593–0.860	0.786	0.720	0.324	0.952	Test
ExtraTrees	0.899	0.915	0.857–0.972	0.815	0.910	0.550	0.973	Training
ExtraTrees	0.760	0.749	0.612–0.887	0.714	0.768	0.345	0.940	Test
LightGBM	0.561	0.738	0.638–0.838	0.815	0.527	0.188	0.955	Training
LightGBM	0.792	0.695	0.531–0.859	0.643	0.817	0.375	0.931	Test

Abbreviations: LR, logistic regression; SVM, support vector machine; LightGBM, light gradient boosting machine; AUC, area under the curve; CI: Confidence Interval; PPV, positive predictive value; NPV, negative predictive value

Fig. 4.

Deep Learning Radiomics (DLR) signature construction and model interpretability. (A) Grad-CAM visualizations for representative lesions with local failure (upper) and local control (lower). (B) LASSO coefficient paths from five-fold cross-validation in the training cohort. (C, D) Comparison of DLR model predictions in the training (C) and test (D) cohorts

To enhance the interpretability of the deep learning model, Gradient-weighted Class Activation Mapping (Grad-CAM) was utilized. This technique generated heatmaps (Fig. 4A) that highlighted image regions most influential for the model’s predictions. The visualizations revealed that the model predominantly focused on peri-lesional areas and intratumoral heterogeneity patterns when distinguishing LC from LF lesions. This analysis provides significant insights into the deep learning framework’s decision-making process.

Comparative evaluation of all five predictive signatures

Table 3 summarizes the performance metrics of the five developed signatures across both training and test cohorts. In the training set, the Combined model had the highest discriminative accuracy, achieving an accuracy of 0.706 and an AUC of 0.854 (95% CI: 0.791–0.917; Fig. 5). The radiomics‑only model closely followed with an accuracy of 0.614 and an AUC of 0.820 (95% CI: 0.745–0.895). Both the DL‑only model (AUC = 0.807, 95% CI: 0.707–0.906) and the DLR model (AUC = 0.783, 95% CI: 0.695–0.870) also exhibited robust predictive capabilities. In contrast, the clinical‑only model showed the lowest discriminative performance among all five models (AUC = 0.746, 95% CI: 0.655–0.837). In the independent test cohort, the Combined model numerically achieved the highest accuracy of 0.792 and an AUC of 0.781 (95% CI: 0.618–0.945). The DLR model also performed favorably in the test set (accuracy = 0.823, AUC = 0.774, 95% CI: 0.630–0.918), ranking second numerically. The radiomics‑only (AUC = 0.748), DL‑only (AUC = 0.733), and clinical‑only (AUC = 0.652) models demonstrated progressively lower predictive accuracy. The DeLong test (Fig. S2) was applied to determine whether the observed differences in AUC values between models reached statistical significance. In the training cohort, the Combined model showed significantly better discriminative ability than the clinical‑only model (p = 0.008) and the DLR model (p = 0.049). However, in the independent test cohort, the Combined model remained significantly superior only to the clinical‑only model (p = 0.024), while comparisons with the other models (radiomics‑only, DL‑only, and DLR) did not reach statistical significance. Despite this, the Combined model consistently showed the highest numerical predictive performance across both cohorts. The integrated multimodal features (Combined and DLR) generally demonstrated improved performance compared to all single-modality models, suggesting the potential added value of multimodal feature fusion. Despite this, the Combined model consistently showed the highest numerical predictive performance across both cohorts. The integrated multimodal features (Combined and DLR) generally demonstrated improved performance compared to all single-modality models, suggesting the potential added value of multimodal feature fusion.

Table 3.

Comparison of five predictive model performance

Signature	Accuracy	AUC	95% CI	Sensitivity	Specificity	PPV	NPV	Cohort
Clinic	0.671	0.746	0.655–0.837	0.778	0.657	0.233	0.957	Training
Rad	0.614	0.820	0.745–0.895	0.963	0.567	0.230	0.991	Training
DL	0.781	0.807	0.707–0.906	0.778	0.781	0.323	0.963	Training
DLR	0.662	0.783	0.695–0.870	0.815	0.642	0.234	0.963	Training
Combined	0.706	0.854	0.791–0.917	0.926	0.677	0.278	0.986	Training
Clinic	0.740	0.652	0.500–0.804	0.643	0.756	0.310	0.925	Test
Rad	0.792	0.748	0.602–0.894	0.643	0.817	0.375	0.931	Test
DL	0.833	0.733	0.583–0.884	0.500	0.890	0.437	0.912	Test
DLR	0.823	0.774	0.630–0.918	0.643	0.854	0.429	0.933	Test
Combined	0.792	0.781	0.618–0.945	0.786	0.793	0.393	0.956	Test

Rad: Radiomics; DL: Deep Learning; DLR: Deep Learning Radiomics; AUC: Area Under Curve; CI: Confidence Interval; PPV: Positive Predictive Value; NPV: Negative Predictive Value

Fig. 5.

Receiver operating characteristic (ROC) curves for the five predictive models (Clinical, Radiomics [Rad], Deep Learning [DL], Deep Learning Radiomics [DLR], and Combined) in the training (A) and the independent test (B) cohorts. AUC: Area Under the Curve

Calibration and clinical utility assessment

The calibration performance of the Combined model was evaluated using the HL goodness-of-fit test, with results visualized through calibration plots (Fig. 6A, B). In the training cohort, the HL test yielded a p value of 0.015, indicating a statistically significant lack of fit, which might suggest minor overfitting. Conversely, the HL test was non-significant in the independent test cohort (p = 0.148; Table 4), implying that the model remained adequately calibrated. Visually, the calibration curves further supported this interpretation, demonstrating reasonable agreement between predicted probabilities and observed outcomes in both cohorts. DCA was conducted to assess the clinical net benefit of each model across various threshold probabilities (Fig. 6C, D). Within the clinically relevant range of 10–50%, the Combined model consistently demonstrated a higher net benefit compared to the DLR model and all single-modality models (clinical-only, radiomics-only, and DL-only) in both the training and test cohorts. These results suggest that using the Combined model for clinical decision-making holds the potential to be more beneficial for patients than the “treat all” or “treat none” strategies across a broad spectrum of clinical scenarios.

Fig. 6.

Calibration and decision curve analysis (DCA) of the five predictive models. (A, B) Calibration curves showing agreement between predicted and observed probabilities in the training (A) and test (B) cohorts. (C, D) Decision curves illustrating the net clinical benefit across threshold probabilities in the training (C) and test (D) cohorts

Table 4.

Hosmer-Lemeshow test

Cohort	Clinic	Rad	DL	DLR	Combined
Training	0.022	0.018	0.071	0.023	0.015
Test	0.012	0.220	0.065	0.067	0.148

Rad: Radiomics; DL: Deep Learning; DLR: Deep Learning Radiomics

To facilitate potential clinical translation, a nomogram derived from the Combined model was developed (Fig. 7). This nomogram visualizes the predicted probability of local failure for individual patients by integrating multimodal imaging signatures and four selected clinical variables: treatment dose, prior brain treatment, prior WBRT, and the presence of extracranial metastases. It provides a practical tool for pretreatment risk stratification. Clinicians can input a patient’s specific clinical and imaging features to obtain an individualized probability estimate of local failure following GKRS, potentially enabling more personalized post-radiosurgical management. However, we acknowledge that the actual impact and utility of this nomogram in real-world clinical decision-making necessitate prospective validation.

Fig. 7.

Nomogram for individualized prediction of local failure risk after GKRS. This nomogram integrates multimodal signatures to estimate the probability of local failure. A predicted probability exceeding the optimal threshold, determined by maximizing the Youden Index (0.579 on the independent test cohort), indicates a high risk of local failure, enabling personalized post-radiosurgical management

Discussion

The most significant finding of this study is the superior discriminatory accuracy and substantial net clinical benefit demonstrated by our Combined model. This outcome directly addresses a critical unmet need in current clinical practice: the lack of validated pre-therapeutic instruments capable of reliably identifying brain metastases at high risk of local failure prior to treatment initiation [19]. While previous studies have established the prognostic value of handcrafted radiomic features [24, 25] and explored deep learning architectures [26] for predicting outcomes in brain metastases, our study advances the field by offering several distinct and incremental contributions through the systematic integration and comparative analysis of these diverse approaches.

Firstly, we performed a comprehensive and rigorous head-to-head comparison of five distinct predictive model architectures (clinical-only, radiomics-only, DL-only, fused DLR, and a fully Combined model) within a single, homogeneous dataset. This systematic benchmarking, specifically tailored for breast and lung cancer brain metastases treated with GKRS, provides robust evidence for the incremental value of multimodal feature fusion over single-modality approaches. Such direct comparative proof is indispensable for guiding future model development beyond isolated modality assessments. Secondly, a pivotal strength of our work lies in the rigorous assessment of clinical utility via DCA. In contrast to many prior studies that primarily report abstract performance metrics (e.g., AUC), our DCA results provide tangible evidence of net clinical benefit across a wide spectrum of clinically relevant probability thresholds (10–50%). This shifts the evaluative focus towards potential real-world clinical consequences and actionable decision support, which is vital for clinical translation. This is further enhanced by our interpretable nomogram, derived from the Combined model, which integrates multimodal imaging signatures with key clinical variables to offer a visual instrument for individualized risk estimation and facilitates a risk stratification framework based on the Youden Index. This allows for categorization into high-risk and low-risk groups, guiding intensified surveillance or standard follow-up protocols, thereby optimizing resource allocation and reducing patient burden. Finally, our focused investigation on local control after GKRS in brain metastases originating from breast and lung cancer addresses a specific knowledge gap. Many existing studies often focus on overall survival or include heterogeneous primary tumor types, which can obscure specific prognostic factors relevant to this particular patient cohort. By narrowing this focus, our study provides more granular insights into the local response patterns of these prevalent metastatic subtypes.

The test AUC of 0.781 achieved by our Combined model is plausibly attributed to the observed multimodal synergy, a finding that resonates with an emerging body of literature advocating hybrid approaches in medical image analysis [15, 27]. The Combined model’s superior net clinical benefit, as evidenced by DCA, underscores its enhanced potential for real-world clinical translation. This analytical approach effectively shifts the evaluative focus from abstract performance metrics (e.g., AUC) toward tangible clinical consequences, enabling the avoidance of unnecessary interventions in low-risk patients while ensuring timely therapeutic escalation for high-risk individuals—advantages not afforded by a uniform “treat-all” or “treat-none” strategy.

Our rigorous DCA findings highlight that comprehensive assessments of clinical utility, including decision-analytic tools, are crucial for predictive models in SRS outcomes, moving beyond traditional measurements. The nomogram, by translating a complex predictive algorithm into an interpretable visual instrument, addresses the frequent criticism of AI systems functioning as “black boxes” and supports their integration into multidisciplinary tumor board discussions [28].

The HL test yielded a non-significant result (p = 0.148) in the independent test cohort, and the accompanying calibration plots demonstrated close concordance between predicted probabilities and observed outcomes. These findings collectively confirm that the nomogram provides well-calibrated risk estimates, thereby strengthening its clinical applicability. While the HL test was statistically significant in the training cohort (p = 0.015), suggesting a modest degree of overfitting, this discrepancy is likely attributable to the inherent optimism in model development using training data. Crucially, this does not undermine the model’s generalizability, as further evidenced by the satisfactory test-set calibration.

While the efficacy of analogous prognostic tools in predicting overall survival among patients with brain metastases has been demonstrated, there remains a paucity of literature specifically addressing their application in predicting local control following focal therapies, such as GKRS [29, 30], particularly within the contexts of breast cancer- and lung cancer-derived metastases. Our study addresses this knowledge gap by offering a multimodal predictive tool for this specific clinical scenario that has been rigorously validated.

A substantial body of research has explored the potential of radiomic-only or DL-only approaches for predicting response to SRS. These investigations have reported AUC values typically ranging from approximately 0.65 to 0.85, with observed variations attributable to differences in cohort size, imaging protocols, and outcome definitions [15, 18]. The test AUC of 0.781 achieved by our Combined model demonstrates a favorable comparison with these prior efforts. Notably, the DCA results provide a level of clinical utility evidence that has been largely absent from earlier studies, and our direct head-to-head comparison of five model architectures (clinical-only, radiomics-only, DL-only, DLR, and Combined) within a single dataset provides a more comprehensive benchmark than previous reports, which typically compared only two or three approaches.

The Grad-CAM visualizations derived from the DL model demonstrated that network predictions are predominantly influenced by peri-lesional regions and intratumoral heterogeneity patterns [31]. These specific regions have been consistently identified as correlating with critical pathological features such as peritumoral edema, microscopic infiltration, and altered vascular permeability. From a radiobiological perspective, the model’s emphasis on the tumor-brain interface is particularly significant, as microscopic infiltration beyond the radiographically visible margin is a primary driver of marginal recurrence following focal SRS [32, 33]. Additionally, the focus on intratumoral heterogeneity likely reflects underlying biological states such as chronic hypoxia or internal necrosis [34]. Hypoxic tumor cells are known to be approximately 2.5 to 3 times more radioresistant than normoxic cells due to the ‘oxygen enhancement effect,’ where the lack of molecular oxygen limits the permanent fixation of radiation-induced DNA damage [35, 36]. These factors have been implicated in radioresistance and early progression following SRS. The enhanced credibility of the present approach is attributed, in part, to this biological plausibility, suggesting that the DL model effectively decodes complex radiographic phenotypes into markers of biological aggressiveness and therapeutic resistance.

The radiomic features ultimately retained following LASSO selection were predominantly from texture categories, such as GLCM and GLRLM, which are known to reflect higher-order spatial heterogeneity [18, 27]. This observation aligns with the established concept that tumor heterogeneity, as captured at the imaging level, often mirrors underlying genetic and microenvironmental diversity, which in turn dictates differential treatment responses. The synergistic and complementary nature of the DL and radiomic features likely accounts for the enhanced performance observed in the DLR and Combined models, particularly when compared to their single-modality counterparts [19]. This dual-modality approach thus leverages distinct yet mutually reinforcing information streams, leading to superior predictive capabilities.

While this study provides significant insights, several limitations warrant consideration. Firstly, the retrospective, single-center design and modest sample size (324 metastases, 41 failure events) inherently carry a risk of selection bias and potential overfitting, especially given the initial high-dimensional feature space. We mitigated this by implementing a rigorous dimensionality reduction pipeline (ICC/correlation filtering, LASSO, and PCA), which reduced the predictors to eight for each modality. We acknowledge that a substantial decrease in DLR model performance was observed between the training (AUC: 0.783) and independent test cohorts (AUC: 0.774), as detailed in Table 3; Fig. 4. This performance drop, while modest for DLR compared to some other models, is a recognized phenomenon in complex models trained on limited event numbers and suggests a degree of overfitting despite our dimensionality reduction efforts. Furthermore, the single-center nature of the data may limit the generalizability of our findings to different clinical environments and imaging protocols. We addressed this by rigorously evaluating all models on an independent test cohort and through 5-fold cross-validation, which provides a more realistic estimate of generalizability. Future studies with larger, multi-institutional datasets are crucial to further enhance model robustness and minimize this discrepancy. Secondly, our lesion-level random splitting strategy, while crucial for capturing response heterogeneity, may introduce within-patient correlation. Specifically, the presence of multiple lesions from the same patient across both training and test sets carries a potential risk of information leakage, which could theoretically lead to an overestimation of model performance. We mitigated this by rigorously balancing baseline characteristics between cohorts (Table 1) to minimize bias. However, we acknowledge that a patient-level validation analysis or the use of mixed-effects models in future work could more rigorously account for this correlation and further strengthen generalizability. Thirdly, the exclusion of patients with > 5 metastases ensured a homogeneous oligometastatic cohort but may limit generalizability to polymetastatic cases. Fourthly, the 2D deep learning approach, while computationally efficient and robust against overfitting in small samples, inherently risks losing 3D volumetric information. Future studies with larger datasets should explore 3D architectures. Fifthly, the analysis relied solely on post-contrast T1-weighted sequences; integrating multiparametric MRI (e.g., T2-FLAIR, DWI) could provide a more comprehensive view of the tumor microenvironment. Sixthly, our study focused exclusively on brain metastases originating from breast and lung cancers. While these represent the most prevalent histologies in clinical practice and ensured a relatively homogeneous cohort for initial model development, the radiobiological diversity across different primary malignancies (e.g., melanoma, renal cell carcinoma) means that the current model’s applicability to other cancer types remains to be established.

Moving forward, the clinical utility of our nomogram requires prospective validation with extended follow-up periods to evaluate longer-term outcomes, such as 12-month local control and progression-free survival, which would better reflect the durability of the treatment response. Integrating imaging signatures with molecular profiling or liquid biopsy remains a promising frontier for developing dynamic [37], multiscale risk-stratification systems in personalized neuro-oncology care.

Conclusions

Our study successfully developed and validated a multimodal predictive signature by integrating clinical parameters, meticulously extracted radiomic features, and 2D deep learning representations. The resulting Combined model consistently demonstrated superior numerical discriminatory performance and conferred the most substantial net clinical benefit within our cohort. These findings underscore its promising potential as a robust tool for predicting local control following GKRS in patients with brain metastases originating from breast or lung cancer. Furthermore, the derived nomogram provides an interpretable and clinically actionable instrument for individualized pre-therapeutic risk assessment. While these results are encouraging, prospective multi-institutional validation remains essential to confirm the model’s real-world impact and ensure its generalizability across diverse clinical settings.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AUC

Area under the curve

BMs

Brain metastases

CNN

Convolutional neural network

DCA

Decision curve analysis

DL

Deep learning

DLR

Deep learning radiomics

GKRS

Gamma Knife radiosurgery

GLCM

Gray‑level co‑occurrence matrix

GLRLM

Gray‑level run‑length matrix

GLSZM

Gray‑level size‑zone matrix

HL

Hosmer-Lemeshow

ICC

Intraclass correlation coefficient

IQR

interquartile range

KPS

Karnofsky Performance Status

LASSO

Least absolute shrinkage and selection operator

LC

Local control

LF

Local failure

LightGBM

Light gradient boosting machine

LR

Logistic regression

MRI

Magnetic resonance imaging

NGTDM

Neighborhood gray‑tone difference matrix

NSCLC

Non-small cell lung cancer

PCA

Principal component analysis

RANO-BM

Response Assessment in Neuro-Oncology for Brain Metastases

ROC

Receiver operating characteristic

ROI

Region of interest

SRS

Stereotactic radiosurgery

SVM

Support vector machine

T1CE

contrast‑enhanced T1‑weighted

WBRT

Whole-brain radiotherapy

Author contributions

MZ: conceptualization, methodology, software, formal analysis, writing original draft; XD: methodology, software, formal analysis; XS: validated the region of interest (ROI) delineations; QC: provided study supervision and article revision; JL: provided study supervision and article revision.

Funding

This work was supported by the National Natural Science Foundation of China (No. 82102830).

Data availability

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

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of the First Affiliated Hospital of Nanjing Medical University (Approval No. 2025-SR-1239) and conducted in accordance with the Declaration of Helsinki. The need for informed consent was waived by the Ethics Committee of the First Affiliated Hospital of Nanjing Medical University due to the retrospective nature of the study and the use of anonymized patient data.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.