Radiomics using contrast-enhanced T1-weighted imaging and clinical features for predicting response to EGFR-TKIs in EGFR-mutated non-small cell lung cancer patients with brain metastases Authors: Lian-Yu Sui, Tian-Ye Zhang, Cheng Cheng, Li-Hong Xing, Huan Meng, Chong Liu, Qi Wang, Jia-Ning Wang, Tian-Shuo Zhang, Kun Liu, Xiao-Ping Yin Published: 2026 ABSTRACT ObjectivesTo construct and validate a model based on clinical characteristics and magnetic resonance imaging (MRI) radiomics to predict 1-year efficacy of epidermal growth factor receptor tyrosine kinase inhibitors (EGFR-TKIs) in patients with EGFR-mutant non-small cell lung cancer (NSCLC) brain metastases (BMs).MethodsThis study retrospectively analyzed data from 338 patients with EGFR-mutant NSCLC BMs from three centers, including MRI, clinical and pathological data, and radiological features. Based on the selected significant radiomic features from intratumoral regions extracted from CE-T1WI, while exploring the value of features in 3/5/8 mm peritumoral regions, seven commonly used machine learning algorithms were compared to select the optimal one for model construction, and the best algorithm was selected for model construction. In the model predicting 1-year therapeutic efficacy, clinical, radiomic, and combined models were constructed separately. The model performance was evaluated using receiver operating characteristic curves.ResultsThe final development cohort comprised 285 patients from Center 1, while the external validation set included 57 patients from Centers 2 and 3. In the model predicting 1-year EGFR-TKIs efficacy, the random forest algorithm, which showed the best application, was used to construct the model. Compared with the radiomic and clinical models, the combined model exhibited superior area under the curve performance in the test set (0.756 vs. 0.644 vs. 0.668). In the external validation set, the combined model achieved an area under the curve of 0.743 (95% CI: 0.604–0.881).ConclusionCompared to single clinical or radiomic models, the combined model was more effective in predicting the 1-year efficacy of EGFR-TKIs in patients with NSCLC BMs with EGFR mutations. Introduction Brain metastases (BMs) are secondary tumors originating from solid organs outside the central nervous system and represent the most common type of brain tumor in adults [1]. During the clinical course, 10–40% of patients with solid tumors develop BMs, and the median survival for patients with synchronous BMs from almost all primary sites does not exceed 12 months [2]. Non-small cell lung cancer (NSCLC) is the primary source of BMs, accounting for ~ 50% of cases [3]. BMs are typically associated with poor prognosis and high mortality in lung cancer. Therefore, precise diagnosis and treatment are important for the clinical management of BMs. Epidermal growth factor receptor (EGFR) mutations are widely recognized as the most common genetic alteration in NSCLC. Compared to NSCLC without BMs, the frequency of EGFR mutations is higher in NSCLC with BMs, and patients with NSCLC harboring EGFR mutations are more prone to developing BMs than those with wild-type EGFR [4]. Given the high sensitivity and specificity of EGFR tyrosine kinase inhibitors (EGFR-TKIs) in treating EGFR-mutated lung cancer, and their increasingly vital role in BM management, these agents are associated with favorable prognosis [5]. EGFR-TKIs therapy has significantly improved survival and life expectancy of patients with EGFR-mutated NSCLC [6, 7]. However, due to tumor heterogeneity and patient-specific factors, there is considerable variability in the sensitivity of NSCLC BMs to EGFR-TKIs. Therefore, personalized and precise diagnosis and treatment of BMs are crucial for improving prognosis and quality of life. Magnetic resonance imaging (MRI) offers the advantage of noninvasively and repeatedly visualizing comprehensive, three-dimensional monitoring for intra-, inter-, and peritumoral information. This effectively compensates for deficiencies such as difficulties in obtaining adequate tissue samples, high costs, and long turnaround times for results. Compared to tissue-based molecular analyses, imaging remains the primary means for noninvasive monitoring of therapeutic responses and surveillance of disease progression [8]. Contrast-enhanced T1-weighted imaging (CE-T1WI) is the preferred conventional method for diagnosing BMs because it enhances the contrast between BMs and surrounding tissues. It can reflect signal variations caused by differences in the marginal regions of BMs and the degree of tumor enhancement, making it more advantageous for capturing lesion heterogeneity [9]. CE-T1WI is the most frequently applied sequencing method for the diagnosis and treatment of BMs [10]. CE-T1WI can sensitively reflect tumor vascular permeability, contrast enhancement degree, and intratumoral heterogeneity, all of which are closely associated with EGFR-TKI sensitivity. Accordingly, CE-T1WI exhibits significant potential in visualizing and quantifying pathophysiological alterations in brain metastases before and after treatment, thus enabling reliable assessment of EGFR-TKI therapeutic efficacy. MRI is an indispensable component of BMs diagnosis and post-treatment monitoring; however, current applications remain at a macroscopic level, primarily limited to subjective observations. In contrast to the limited information provided by traditional MRI diagnostic methods, MRI radiomics can automatically quantify tumor heterogeneity, thereby improving the accuracy of diagnostic and prognostic models [11, 12]. The visualization of tumor heterogeneity also aids physicians in formulating more precise treatment plans. Therefore, by studying characteristic changes in tumors on MR images, radiomics methods can predict tumor sensitivity to molecular therapies. Such techniques contribute to making more optimized clinical decisions without incurring additional costs [13]. There is a high incidence of EGFR-mutated NSCLC with BMs, and BMs are an indicator of therapeutic efficacy; therefore, MRI of BMs has the potential to identify novel biomarkers for predicting response to EGFR-TKIs. Previous studies, in predicting the therapeutic efficacy of EGFR-TKIs, have primarily relied on constructing models using radiomic features alone, overlooking the potential complementary value of clinical variables. Integration of clinical characteristics with radiomic analysis can enhance the robustness of models and facilitate exploration of their clinical translational value [14]. The application of MRI radiomics, a noninvasive, quantifiable, and reproducible intelligent method, to evaluate the therapeutic efficacy of malignant tumors in clinical settings has become an inevitable trend. Predictive models constructed based on radiomic features and clinical characteristics demonstrate significant promise in forecasting therapeutic outcomes. The incorporation of radiomic features into clinical models can enhance prediction, thereby improving clinical decision-making without incurring additional costs. This approach is conducive to optimizing treatment plans tailored to the individual needs and risk profiles of patients [15, 16]. Most previous radiomics studies were single‑center, used only radiomic features, and relied on a single MRI sequence. Therefore, the present study extracted radiomic features from MRI scans of patients with EGFR-mutated NSCLC and BMs, integrated multiple data sources, and developed and externally validated a radiomic framework in conjunction with clinical characteristics to predict the therapeutic response to EGFR-TKIs. This study also showed the potential clinical value of the combined model in guiding personalized treatment strategies. Materials and methods Inclusion and exclusion criteria of patients This was a retrospective analysis of brain MRI data from 285 patients treated at Center 1 from April 2016 to January 2024, who served as the primary cohort. For dataset partitioning, the hold‑out method was employed. All enrolled samples were randomly divided into independent training and internal test sets at a ratio of 8:2 using stratified random sampling. Notably, the internal test set was strictly excluded from all subsequent feature selection procedures to ensure unbiased model evaluation. To ensure the representativeness of data partitioning and avoid model bias caused by uneven sample distribution, we stratified the sampling based on efficacy groups. This stratified sampling strategy effectively maintained a highly consistent proportion of the key outcome variable between the training and test sets, while retaining completely independent data that were not involved in model training for the objective validation and evaluation of the model’s true generalization ability. The training and internal test sets of the primary cohort were allocated in an 8:2 ratio for model development and validation. Additionally, 53 NSCLC patients with BMs treated at Center 2 from June 2020 to January 2024 and Center 3 from November 2021 to March 2024 were combined into a new dataset as the external validation set to balance the distribution of demographic and clinicopathological factors. The Institutional Review Board waived the requirement for written consent because of the retrospective nature of the study and the absence of identifiable patient information. Patients with EGFR-mutated NSCLC received EGFR-TKIs therapy in accordance with current clinical guidelines. For each patient, the response to EGFR-TKI treatment was assessed using the response evaluation criteria in solid tumors (RECIST) version 1.1 [17]. Clinical characteristics, including age, gender, EGFR mutation type, EGFR-TKIs type, extracranial metastasis, number of BMs, size of BMs, and peritumoral edema, were obtained from clinical and pathological records, as well as baseline cranial MRI images. The therapeutic efficacy of EGFR-TKIs was determined by evaluating changes in the patients’ condition within 1 year after initiation of EGFR-TKIs treatment. The primary study endpoint was disease status at one year following the initiation of EGFR‑TKI treatment. Intracranial efficacy was assessed using the RECIST criteria, categorized as intracranial complete response (iCR), intracranial partial response (iPR), intracranial stable disease (iSD), and intracranial progressive disease (iPD). The intracranial objective response rate represented the proportion of patients achieving iCR and iPR, while the intracranial disease control rate represented the proportion of patients achieving iCR, iPR, and iSD. Patients with iCR and iPR were classified into the effective treatment group, whereas those with iSD and iPD were classified into the ineffective treatment group. The inclusion criteria were: (1) definitive diagnosis of NSCLC was established through lung biopsy or open surgery; (2) enhanced brain MRI was performed at our hospital within 2 weeks prior to treatment; (3) brain MRI revealed at least one BM; (4) EGFR mutation was confirmed by genetic testing, and the patient received EGFR-TKI treatment; and (5) complete medical records were available, including treatment information and monitoring examination results. The exclusion criteria were: (1) patients did not receive EGFR-TKIs after onset of BMs; (2) poor image quality affected quantitative analysis; (3) incomplete clinical information; (4) other concurrent cancers; and (5) patients who simultaneously received local treatment, such as whole-brain radiotherapy and surgical resection. For more detailed information on the inclusion and exclusion criteria for NSCLC BMs in this study, please refer to Fig. 1. Fig. 1Flowchart of patient inclusion and exclusion criteria. BMs, brain metastases; EGFR, epidermal growth factor receptor; EGFR-TKIs, EGFR tyrosine kinase inhibitors; NSCLC, non-small cell lung cancer Flowchart of patient inclusion and exclusion criteria. BMs, brain metastases; EGFR, epidermal growth factor receptor; EGFR-TKIs, EGFR tyrosine kinase inhibitors; NSCLC, non-small cell lung cancer MRI All three research centers utilized 1.5 T or 3 T MR scanners to perform cranial MRI. After administering a weight-adjusted dose of 0.1 mmol/kg gadobenate dimeglumine contrast agent via an elbow vein using a high-pressure injector at an injection rate of 2 ml/s, axial CE-T1WI sequences were acquired. The imaging protocols varied among the different scanners at each research center. Detailed scanning parameters are provided in Table 1. Table 1CE-T1WI parameters of different MRI scannersScannerField StrengthTR/TE (ms)MatrixFOV (mm)Slice gap (mm)Section thickness (mm)NEXUnited Imaging uMR 5881.5 T450/12.5256 × 75230 × 2001.55.01GE Discovery MR7503.0 T1850/24256 × 320240 × 2401.55.01Siemens Amira1.5 T1640/10256 × 189230 × 2301.55.01Philips Achieva3.0 T120/1.91256 × 164230 × 1831.06.01Abbreviations: CE-T1WI, contrast-enhanced T1 weighted imaging; MRI, magnetic resonance imaging; TR, repetition time; TE, echo time; FOV, field of view; NEX, number of excitations Table 1CE-T1WI parameters of different MRI scannersScannerField StrengthTR/TE (ms)MatrixFOV (mm)Slice gap (mm)Section thickness (mm)NEXUnited Imaging uMR 5881.5 T450/12.5256 × 75230 × 2001.55.01GE Discovery MR7503.0 T1850/24256 × 320240 × 2401.55.01Siemens Amira1.5 T1640/10256 × 189230 × 2301.55.01Philips Achieva3.0 T120/1.91256 × 164230 × 1831.06.01Abbreviations: CE-T1WI, contrast-enhanced T1 weighted imaging; MRI, magnetic resonance imaging; TR, repetition time; TE, echo time; FOV, field of view; NEX, number of excitations CE-T1WI parameters of different MRI scanners Abbreviations: CE-T1WI, contrast-enhanced T1 weighted imaging; MRI, magnetic resonance imaging; TR, repetition time; TE, echo time; FOV, field of view; NEX, number of excitations Radiomics analysis Figure 2 illustrates the specific workflow of radiomic analysis. To minimize non-biological technical variations introduced by differences in medical centers, MRI scanners (1.5 T vs. 3.0 T), and scanning protocols, Z-score normalization was applied to all radiomic features for consistency correction prior to feature dimensionality reduction. Specifically, each feature value was scaled to a standard distribution with a mean of 0 and a variance of 1 by subtracting its mean and dividing by its standard deviation. This procedure effectively reduced baseline signal discrepancies across different scanners, ensuring that multicenter features were processed on a unified scale in the subsequent workflows of radiomic feature extraction, selection, and machine learning model construction. Fig. 2Flowchart of the proposed model analyses. DT, decision tree; GLCM, gray-level co-occurrence matrix; GLDM, gray-level dependence matrix; GLRLM, gray-level run-length matrix; GLSZM, gray-level size zone matrix; GNB, Gaussian naive Bayes; KNN, k-nearest neighbors; LASSO, least absolute shrinkage and selection operator; NGTDM, Neighboring gray-tone difference matrix; LR, logistic regression; MRI, magnetic resonance imaging; RF, random forest; SVM, support vector machine; XGB, extreme gradient boosting Flowchart of the proposed model analyses. DT, decision tree; GLCM, gray-level co-occurrence matrix; GLDM, gray-level dependence matrix; GLRLM, gray-level run-length matrix; GLSZM, gray-level size zone matrix; GNB, Gaussian naive Bayes; KNN, k-nearest neighbors; LASSO, least absolute shrinkage and selection operator; NGTDM, Neighboring gray-tone difference matrix; LR, logistic regression; MRI, magnetic resonance imaging; RF, random forest; SVM, support vector machine; XGB, extreme gradient boosting MRI segmentation All imaging data were retrieved from the Picture Archiving and Communication System and stored in DICOM format. Two radiologists used ITK-SNAP version 3.8.0 (www.itksnap.org) to delineate regions of interest (ROIs) on axial CE-T1WI scans. The two radiologists involved in ROI segmentation have 5 years and 10 years of experience in neuroimaging diagnosis, respectively, and both have received standardized training in radiomics ROI delineation. A computer-generated random number table was used to select images from 30 patients for intra- and interobserver consistency analysis. The imaging data of these 30 patients were used for tumor ROI delineation by two radiologists who were blinded to the patients’ pathological results. One of the radiologists repeated the tumor ROI delineation 1 month after the initial delineation. Interobserver consistency analysis was based on the data from the initial delineation by both radiologists, while the intraclass correlation coefficient (ICC) was conducted using the data from the two delineations performed by the same radiologist (with 10 years of experience). Radiomic feature extraction and selection The original images underwent N4_Normalization processing, and radiomic features were extracted from the ROIs of each patient using Pyradiomics version 3.1.0. The radiomic features extracted from each ROI included: morphology (e.g., volume, surface area, and diameter); first-order statistical features (e.g., mean, standard deviation, and entropy). Based on first-order features (which only consider the gray-level distribution of individual pixels), this study extracted a variety of second-order features (i.e., texture features). Second-order features are mainly used to describe the spatial relationships and arrangement patterns among voxels in images, and can reflect the intratumoral micro-heterogeneity at a deeper level. The second-order features included: gray-level co-occurrence matrix (describing gray-level spatial correlation); gray-level run-length matrix (analyzing texture roughness); gray-level size zone matrix (quantifying distribution of region sizes); gray-level dependence matrix (describing local gray-level variations); and neighborhood gray-tone difference matrix (analyzing local heterogeneity). For the selection of radiomic features, an ICC consistency test was first performed, and features with an ICC value greater than 0.8 were retained. A total of 1505 radiomic features were extracted from each patient; after excluding 430 features with an ICC value less than 0.8, 1075 highly reproducible features remained for subsequent analysis. The median ICC of the retained features was 0.974, with an interquartile range of 0.081 (0.913–0.994) and a value range of 0.802–1.000. Subsequently, Spearman correlation analysis was applied for the initial dimensionality reduction of radiomic features, and features with an absolute correlation coefficient greater than 0.9 were excluded. This step eliminated 835 highly redundant features, yielding 240 robust and independent radiomic features. A recursive feature elimination algorithm was then used to iteratively remove unimportant features. Finally, based solely on the training subset, the least absolute shrinkage and selection operator (LASSO) regression combined with 10-fold cross-validation was adopted to complete the final feature dimensionality reduction and screen out the optimal feature set. This cross-validation strategy was exclusively used for feature optimization, instead of model training and testing. Construction and validation of predictive models Model training Different combinations of feature types and classifiers can be customized to suit specific classification tasks. The motivation for using multiple classifiers was to enhance the robustness of the results, as previous studies have indicated that selecting an appropriate algorithm plays a crucial role in classification outcomes [18, 19]. This study recognized that no single model in machine learning is universally applicable and explored various classifiers to develop a high-performance and robust model for predicting the therapeutic efficacy of EGFR-TKIs in NSCLC BMs. To adhere to best ML practices, a combined classifier comprising seven commonly used ML algorithms was adopted, comprising: logistic regression as a linear model; random forest (RF) as an ensemble method; k-nearest neighbors for instance-based modeling; linear support vector machine; decision tree as a tree-structured supervised learning model; Gaussian naïve Bayes as a probability classification model; and extreme gradient boosting as a gradient boosting framework. In this study, machine learning models were constructed and selected using a rigorous evaluation procedure. First, for hyperparameter optimization, parameter tuning was performed for each candidate machine learning algorithm on the training set. Grid search was applied to optimize key model parameters, such as the regularization type (L1 or L2) and regularization strength (e.g., C value) for logistic regression, as well as the number of trees and maximum depth for random forest, in order to identify the parameter combination that achieved optimal performance on the training data without overfitting. Furthermore, to objectively evaluate the generalization ability of the models and select the final algorithm, each trained model was directly applied to the independent test set that was not involved in model training. Subsequently, the seven ML models were evaluated, and their efficacies were comprehensively assessed through area under the curve (AUC) analysis, decision curve analysis to determine standardized net benefit, and calibration curves for each trained model to identify the optimal model. After performing feature extraction and selection from within the BMs, as well as from automatically expanded 3-, 5-, and 8-mm peritumoral regions, radiomic models were constructed based on optimal ML algorithms according to model training, and their performances were compared. Predictive model construction Clinical and pathological data were retrospectively collected from electronic medical records and digital pathology systems across various centers, including: age; gender; smoking status; Karnofsky Performance Status score; primary tumor site; pathological classification; tumor markers [carcinoembryonic antigen, neuron-specific enolase, cytokeratin 19 fragment antigen 21‑1 (CYFRA 21‑1)]; extracranial metastasis; TNM staging (describing the size and invasiveness of the primary tumor, lymph node metastasis, and systemic metastasis); EGFR mutation subtype; and EGFR-TKIs type. Additionally, MRI morphological features commonly used in clinical assessments of central nervous system tumors were analyzed, including; number and grouping of BMs; location, maximum diameter, and grouping of the largest lesion; enhancement pattern; and peritumoral edema. Logistic regression analysis was conducted to select clinically significant features (P < 0.05) for model construction. Finally, radiomic, clinical, and combined models were constructed based on the optimal ML algorithm. The optimal model was determined by comparing the AUCs using the DeLong test, and the model’s generalization capability was further evaluated in an external validation set. Statistical analysis We used SPSS 22.0 software (IBM Corporation, Chicago, IL, USA) for statistical analysis. In descriptive statistics, the normality of continuous variables was tested using the Shapiro–Wilk test. Continuous variables were expressed as mean ± standard deviation or median (interquartile range), while categorical variables were presented as frequencies and percentages. For analysis of differences among the training, test, and external validation sets, continuous variables were compared using Student’s t test or Mann–Whitney U test, depending on the data type, while categorical variables were compared using χ2 test or Fisher’s exact test. The discriminatory ability of predictive models was evaluated and compared using the 95% confidence interval (CI) of the area under the receiver operating characteristic curve (ROC). Model performance was further assessed by calculating AUC, accuracy (ACC), sensitivity (SEN), specificity (SPE), and F1 score. The Delong test was used to compare the AUC values of different models and evaluate their robustness. The performance of radiomic, clinical, and combined models was evaluated separately in the test and external validation sets. Calibration curves were used to assess the deviation in the distribution of predicted values, reflecting the consistency between model predictions and actual probabilities. Based on this, decision curve analysis was conducted to calculate the net benefit at different threshold probabilities, thereby evaluating the model’s value in differential diagnosis. All statistical analyses were performed with a significance level set at P < 0.05. Results Patient’s basic information and treatment efficacy A total of 338 patients from three centers were included. Table 2 summarizes the clinical characteristics, Table 3 the MRI features, and Table 4 the therapeutic efficacy of EGFR-TKIs. The clinical features included KPS score, extracranial metastases, M staging of the TNM classification, and clinical staging. The MRI morphological features included the number of lesions, necrosis, and EGFR-TKIs type. Significant differences were observed across the training, test, and external validation sets (all P < 0.05). The assessment showed that 39 patients achieved iCR, 104 iPR, 60 iSD, and 131 iPD. The intracranial objective response rate was 42.3%, and the intracranial disease control rate was 60.1%. Table 2Clinical characteristics of NSCLC BMs patientsClinical characteristicsTraining set (N = 228) Test set (N = 57)External validation set (N = 53)Overall (N = 338)P valueAge (M ± SD, years)59.8 ± 9.562.0 ± 9.859.5 ± 8.761.0 ± 10.20.345Gender (N, %)Male51 (22.4)13(22.8)11(20.8)84 (24.9)0.061 Female177 (77.6)44(77.2)33(62.3)254 (75.1)Tobacco use N, %)Non-smokers166 (72.8)37(74.9)37(69.8)240 (71.0)0.491Smoker62 (27.2)20(35.1)16(30.2)98 (29.0)Basic disease* (N, %)No125 (54.8)28(49.1)34(64.2)187 (55.3)0.275Yes103 (45.2)29(50.9)19(35.8)151 (44.7)KPS score (N, %)≤ 7049 (21.5)14(24.6)31(58.5)94 (27.8) < 0.001 > 70179 (78.5)43(75.4)22(41.5)244 (72.2)Extracranial metastases (N, %)No68 (29.8)19(33.3)27(50.9)114 (33.7) 0.014 Yes160 (70.2)38(66.7)26(49.1)224 (66.3)Location of LC (N, %)Left lung89 (39.0)23(40.4)19(35.8)131(38.8)0.879Right lung139 (61.0)34(59.6)34(64.2)207 (61.2)Lung surgery (N, %)No184 (80.7)47(82.5)47(88.7)278 (82.2)0.391Yes44 (19.3)10(17.5)6(11.3)60 (17.8)T staging of TNM (N, %)T1 + T299(56.9)26(63.4)15(60.0)140 (58.3)0.736T3 + T475(43.1)15(36.6)10(40.0)100 (41.7)N staging TNM (N, %)N0 + N155 (31.4)11(25.6)7(25.9)73 (29.8)0.677N2 + N3120 (68.6)32(74.4)20(74.1)172 (70.2)M staging of TNM (N, %)M1b27 (11.8)9(15.8)27(50.9)63 (18.6) < 0.001 M1c141 (18.0)10(17.5)18(34.0)69 (20.4)M1c2160 (70.2)38(66.7)8(15.1)206 (60.9)Clinical staging (N, %)Ⅳ A27 (11.8)9(15.8)22(41.5)58 (17.2) < 0.001 Ⅳ B201 (88.2)48(84.2)31(58.5)280 (82.8)Pathology type (N, %)AD221(96.9)57(100)52(98.1)330 (97.6)0.165SCC1 (0.4)01(1.9)2 (0.6)Others6 (2.6)006 (1.8)CEA (N, %)≤ 4.7 ng/ml63 (27.6)13 (22.8)16 (30.2)92 (27.2)0.665> 4.7 ng/ml165 (72.4)44 (77.2)37 (69.8)246 (72.8)NSE (N, %)≤ 16.3 ng/ml106 (46.5)27 (47.4)18 (34.0)151 (44.7)0.231> 16.3 ng/ml122 (53.5)30 (52.6)35 (66.0)187 (55.3)CYFRA 21‑1 (N, %)≤ 3.3 ng/ml62 (27.2)18 (31.6)23 (43.4)103 (30.5)0.068> 3.3 ng/ml166 (72.8)39 (68.4)30 (56.6)235 (69.5)*Basic diseases: diabetes, hypertension, and coronary heart disease. AD, adenocarcinoma; BMs, brain metastasis; CEA, carcinoembryonic antigen; CYFRA 21‑1, cytokeratin 19 fragment antigen 21‑1; NSE, neuron-specific enolase; SCC, squamous cell cancer; NSCLC, non-small cell lung cancer. P value is derived from the c2 test or Kruskal-Wallis test for each variable among the training, test and external validation sets. Bold text indicates P < 0.05 Table 2Clinical characteristics of NSCLC BMs patientsClinical characteristicsTraining set (N = 228) Test set (N = 57)External validation set (N = 53)Overall (N = 338)P valueAge (M ± SD, years)59.8 ± 9.562.0 ± 9.859.5 ± 8.761.0 ± 10.20.345Gender (N, %)Male51 (22.4)13(22.8)11(20.8)84 (24.9)0.061 Female177 (77.6)44(77.2)33(62.3)254 (75.1)Tobacco use N, %)Non-smokers166 (72.8)37(74.9)37(69.8)240 (71.0)0.491Smoker62 (27.2)20(35.1)16(30.2)98 (29.0)Basic disease* (N, %)No125 (54.8)28(49.1)34(64.2)187 (55.3)0.275Yes103 (45.2)29(50.9)19(35.8)151 (44.7)KPS score (N, %)≤ 7049 (21.5)14(24.6)31(58.5)94 (27.8) < 0.001 > 70179 (78.5)43(75.4)22(41.5)244 (72.2)Extracranial metastases (N, %)No68 (29.8)19(33.3)27(50.9)114 (33.7) 0.014 Yes160 (70.2)38(66.7)26(49.1)224 (66.3)Location of LC (N, %)Left lung89 (39.0)23(40.4)19(35.8)131(38.8)0.879Right lung139 (61.0)34(59.6)34(64.2)207 (61.2)Lung surgery (N, %)No184 (80.7)47(82.5)47(88.7)278 (82.2)0.391Yes44 (19.3)10(17.5)6(11.3)60 (17.8)T staging of TNM (N, %)T1 + T299(56.9)26(63.4)15(60.0)140 (58.3)0.736T3 + T475(43.1)15(36.6)10(40.0)100 (41.7)N staging TNM (N, %)N0 + N155 (31.4)11(25.6)7(25.9)73 (29.8)0.677N2 + N3120 (68.6)32(74.4)20(74.1)172 (70.2)M staging of TNM (N, %)M1b27 (11.8)9(15.8)27(50.9)63 (18.6) < 0.001 M1c141 (18.0)10(17.5)18(34.0)69 (20.4)M1c2160 (70.2)38(66.7)8(15.1)206 (60.9)Clinical staging (N, %)Ⅳ A27 (11.8)9(15.8)22(41.5)58 (17.2) < 0.001 Ⅳ B201 (88.2)48(84.2)31(58.5)280 (82.8)Pathology type (N, %)AD221(96.9)57(100)52(98.1)330 (97.6)0.165SCC1 (0.4)01(1.9)2 (0.6)Others6 (2.6)006 (1.8)CEA (N, %)≤ 4.7 ng/ml63 (27.6)13 (22.8)16 (30.2)92 (27.2)0.665> 4.7 ng/ml165 (72.4)44 (77.2)37 (69.8)246 (72.8)NSE (N, %)≤ 16.3 ng/ml106 (46.5)27 (47.4)18 (34.0)151 (44.7)0.231> 16.3 ng/ml122 (53.5)30 (52.6)35 (66.0)187 (55.3)CYFRA 21‑1 (N, %)≤ 3.3 ng/ml62 (27.2)18 (31.6)23 (43.4)103 (30.5)0.068> 3.3 ng/ml166 (72.8)39 (68.4)30 (56.6)235 (69.5)*Basic diseases: diabetes, hypertension, and coronary heart disease. AD, adenocarcinoma; BMs, brain metastasis; CEA, carcinoembryonic antigen; CYFRA 21‑1, cytokeratin 19 fragment antigen 21‑1; NSE, neuron-specific enolase; SCC, squamous cell cancer; NSCLC, non-small cell lung cancer. P value is derived from the c2 test or Kruskal-Wallis test for each variable among the training, test and external validation sets. Bold text indicates P < 0.05 Clinical characteristics of NSCLC BMs patients *Basic diseases: diabetes, hypertension, and coronary heart disease. AD, adenocarcinoma; BMs, brain metastasis; CEA, carcinoembryonic antigen; CYFRA 21‑1, cytokeratin 19 fragment antigen 21‑1; NSE, neuron-specific enolase; SCC, squamous cell cancer; NSCLC, non-small cell lung cancer. P value is derived from the c2 test or Kruskal-Wallis test for each variable among the training, test and external validation sets. Bold text indicates P < 0.05 Table 3MRI radiological characteristics of NSCLC BMs patientsRadiological characteristicsTraining set (N = 228)Test set (N = 57)External validation set (N = 53)Overall (N = 338)P valueNumber of lesions (N, %) < 0.001 1–249 (21.5)13 (22.8)29 (54.7)91 (26.9)3–10141 (61.8)35 (61.4)19 (35.8)195 (57.7)> 1038 (16.7)9 (15.8)5 (9.4)52 (15.4)Size (N, %)0.060≤ 0.5 cm41 (18.0)5 (8.8)9 (17.0)55 (16.3)0.5 cm -3 cm177 (77.6)50 (18.9)37 (69.8)264 (78.1)> 3 cm10 (4.4)2 (3.5)7 (13.2)19 (5.6)Enhancement patterns (N, %)0.350Ring hyperintensity49 (21.5)16 (28.1)10 (18.9)75 (22.2)Homogeneous hyperintensity139 (61.0)35 (61.4)30 (56.6)204 (60.4)Hypointensity / isointensity40 (17.5)6 (10.5)13 (24.5)59 (17.5)Peritumoral edema (N, %)0.492Yes132 (57.9)33 (57.9)26 (49.1)191 (56.5)No96 (42.1)24 (42.1)27 (50.9)147 (43.5)Necrosis (N, %) 0.032 Yes155 (68.0)38 (66.7)26 (49.1)219 (64,8)No73 (32.0)19 (33.3)27 (50.9)119 (35.2)Location of BMs (N, %)0.467Supratentorial area119 (52.2)32 (56.1)29 (54.7)180 (53.3)Infratentorial area13 (5.7)2 (3.5)6 (11.3)21 (6.2)Supra- and infratentorial area96 (42.1)23 (40.4)18 (34.0)137 (40.5)EGFR-mutant type (N, %)0.52319 exon110 (48.2)27 (47.4)21 (39.6)158 (46.7)21exon118 (51.8)30 (52.6)32 (60.4)180 (53.3)EGFR-TKIs type* (N, %) 0.005 First generation143 (62.7)33 (57.9)22 (41.5)198 (58.6)Second generation6 (2.6)4 (7.0)010 (3.0)Third generation79 (34.6)20 (35.1)31 (58.5)130 (38.5)*EGFR‑TKIs were categorized as follows: first‑generation (gefitinib, erlotinib, icotinib); second‑generation (afatinib, dacomitinib); third‑generation (osimertinib, almonertinib, furmonertinib). BMs, brain metastasis; EGFR, epidermal growth factor receptor; EGFR-TKIs, EGFR tyrosine kinase inhibitors; NSCLC, non-small cell lung cancer. P value is derived from the c2 test or Kruskal-Wallis test for each variable among the training, test and external validation sets. Bold text indicates P < 0.05 Table 3MRI radiological characteristics of NSCLC BMs patientsRadiological characteristicsTraining set (N = 228)Test set (N = 57)External validation set (N = 53)Overall (N = 338)P valueNumber of lesions (N, %) < 0.001 1–249 (21.5)13 (22.8)29 (54.7)91 (26.9)3–10141 (61.8)35 (61.4)19 (35.8)195 (57.7)> 1038 (16.7)9 (15.8)5 (9.4)52 (15.4)Size (N, %)0.060≤ 0.5 cm41 (18.0)5 (8.8)9 (17.0)55 (16.3)0.5 cm -3 cm177 (77.6)50 (18.9)37 (69.8)264 (78.1)> 3 cm10 (4.4)2 (3.5)7 (13.2)19 (5.6)Enhancement patterns (N, %)0.350Ring hyperintensity49 (21.5)16 (28.1)10 (18.9)75 (22.2)Homogeneous hyperintensity139 (61.0)35 (61.4)30 (56.6)204 (60.4)Hypointensity / isointensity40 (17.5)6 (10.5)13 (24.5)59 (17.5)Peritumoral edema (N, %)0.492Yes132 (57.9)33 (57.9)26 (49.1)191 (56.5)No96 (42.1)24 (42.1)27 (50.9)147 (43.5)Necrosis (N, %) 0.032 Yes155 (68.0)38 (66.7)26 (49.1)219 (64,8)No73 (32.0)19 (33.3)27 (50.9)119 (35.2)Location of BMs (N, %)0.467Supratentorial area119 (52.2)32 (56.1)29 (54.7)180 (53.3)Infratentorial area13 (5.7)2 (3.5)6 (11.3)21 (6.2)Supra- and infratentorial area96 (42.1)23 (40.4)18 (34.0)137 (40.5)EGFR-mutant type (N, %)0.52319 exon110 (48.2)27 (47.4)21 (39.6)158 (46.7)21exon118 (51.8)30 (52.6)32 (60.4)180 (53.3)EGFR-TKIs type* (N, %) 0.005 First generation143 (62.7)33 (57.9)22 (41.5)198 (58.6)Second generation6 (2.6)4 (7.0)010 (3.0)Third generation79 (34.6)20 (35.1)31 (58.5)130 (38.5)*EGFR‑TKIs were categorized as follows: first‑generation (gefitinib, erlotinib, icotinib); second‑generation (afatinib, dacomitinib); third‑generation (osimertinib, almonertinib, furmonertinib). BMs, brain metastasis; EGFR, epidermal growth factor receptor; EGFR-TKIs, EGFR tyrosine kinase inhibitors; NSCLC, non-small cell lung cancer. P value is derived from the c2 test or Kruskal-Wallis test for each variable among the training, test and external validation sets. Bold text indicates P < 0.05 MRI radiological characteristics of NSCLC BMs patients *EGFR‑TKIs were categorized as follows: first‑generation (gefitinib, erlotinib, icotinib); second‑generation (afatinib, dacomitinib); third‑generation (osimertinib, almonertinib, furmonertinib). BMs, brain metastasis; EGFR, epidermal growth factor receptor; EGFR-TKIs, EGFR tyrosine kinase inhibitors; NSCLC, non-small cell lung cancer. P value is derived from the c2 test or Kruskal-Wallis test for each variable among the training, test and external validation sets. Bold text indicates P < 0.05 Table 4Intracranial efficacy of EGFR-TKIs in NSCLC BMs patientsEfficiencyTraining set (N = 228)Test set (N = 57)External validation set (N = 53)Overall (N = 338)iCR (N, %)30 (13.2)7 (12.3)2 (3.8)39 (11.5)iPR (N, %)68 (29.8)18 (31.6)22 (41.5)108 (32.0)iSD (N, %)42 (18.4)8 (14.0)10 (18.9)60 (17.8)iPD (N, %)88 (38.6)24 (42.1)19 (35.8)131 (38.8)iORR43.0%43.9%45.3%43.5%iDCR61.4%57.9%64.2%61.2%iCR, intracranial complete response; iPR, intracranial partial response; iSD, intracranial stable disease; iPD, intracranial progressive disease; BMs, brain metastasis; EGFR, epidermal growth factor receptor; EGFR-TKIs, EGFR tyrosine kinase inhibitors; NSCLC, non-small cell lung cancer Table 4Intracranial efficacy of EGFR-TKIs in NSCLC BMs patientsEfficiencyTraining set (N = 228)Test set (N = 57)External validation set (N = 53)Overall (N = 338)iCR (N, %)30 (13.2)7 (12.3)2 (3.8)39 (11.5)iPR (N, %)68 (29.8)18 (31.6)22 (41.5)108 (32.0)iSD (N, %)42 (18.4)8 (14.0)10 (18.9)60 (17.8)iPD (N, %)88 (38.6)24 (42.1)19 (35.8)131 (38.8)iORR43.0%43.9%45.3%43.5%iDCR61.4%57.9%64.2%61.2%iCR, intracranial complete response; iPR, intracranial partial response; iSD, intracranial stable disease; iPD, intracranial progressive disease; BMs, brain metastasis; EGFR, epidermal growth factor receptor; EGFR-TKIs, EGFR tyrosine kinase inhibitors; NSCLC, non-small cell lung cancer Intracranial efficacy of EGFR-TKIs in NSCLC BMs patients iCR, intracranial complete response; iPR, intracranial partial response; iSD, intracranial stable disease; iPD, intracranial progressive disease; BMs, brain metastasis; EGFR, epidermal growth factor receptor; EGFR-TKIs, EGFR tyrosine kinase inhibitors; NSCLC, non-small cell lung cancer Predictive models evaluation A higher AUC indicated greater classification accuracy, while a smaller AUC gap between the training and test sets suggested stronger model generalizability and lower risk of overfitting. The model training results (Fig. 3) indicated that the radiomic model constructed based on the RF algorithm emerged as the top performer. Comparison of RF models derived from intratumoral features of BMs and extended 3-, 5-, and 8-mm peritumoral regions revealed that intratumoral features demonstrated better AUC performance in the test set (Fig. 4). Therefore, intratumoral features were selected for constructing the radiomic model. LASSO was applied to determine the optimal parameter value (λ = 0.0285) based on intratumoral features of BMs, ultimately selecting eight significant features that made important contributions to the classification of the predictive model (Fig. 5). The detailed information of these 8 radiomic features is summarized in Table 5. These 8 features were used to calculate a radiomics score (Rad-score). The radiomic feature finally input for model training was the Rad-score, computed as: Fig. 3ML Model training. ROC curve analysis and AUC (a, b), decision curve analysis (DCA) (c, d), and calibration curve for the validation set (e) of the classifier models in the training and test sets. The ROC curve describes the performance of the classifier models in predicting treatment efficacy. The AUC values for the training and test sets are provided to evaluate the discriminatory ability of the models. A higher AUC value indicates greater classification accuracy, while a smaller gap in AUC between the training and validation sets suggests stronger model generalization ability and a lower risk of overfitting. DCA directly calculates the practical value of the models under different decision thresholds. The calibration curve assesses the degree of match between the predicted probabilities and the actual probabilities for each model. AUC, area under the curve; DT, decision tree; GNB, Gaussian naive Bayes; KNN, k-nearest neighbors; LR, logistic regression; ML, machine learning; RF, random forest; SVM, support vector machine; XGB, extreme gradient boosting ML Model training. ROC curve analysis and AUC (a, b), decision curve analysis (DCA) (c, d), and calibration curve for the validation set (e) of the classifier models in the training and test sets. The ROC curve describes the performance of the classifier models in predicting treatment efficacy. The AUC values for the training and test sets are provided to evaluate the discriminatory ability of the models. A higher AUC value indicates greater classification accuracy, while a smaller gap in AUC between the training and validation sets suggests stronger model generalization ability and a lower risk of overfitting. DCA directly calculates the practical value of the models under different decision thresholds. The calibration curve assesses the degree of match between the predicted probabilities and the actual probabilities for each model. AUC, area under the curve; DT, decision tree; GNB, Gaussian naive Bayes; KNN, k-nearest neighbors; LR, logistic regression; ML, machine learning; RF, random forest; SVM, support vector machine; XGB, extreme gradient boosting Fig. 4RF-based radiomics prediction models for intratumoral and peritumoral (3/5/8 mm) regions. ROC curve and AUC (a) and decision curve (b) in the test set. AUC, area under the curve; RF, random forest; ROC, receiver operating characteristic RF-based radiomics prediction models for intratumoral and peritumoral (3/5/8 mm) regions. ROC curve and AUC (a) and decision curve (b) in the test set. AUC, area under the curve; RF, random forest; ROC, receiver operating characteristic Fig. 5Feature selection using 10-fold cross-validation LASSO. Eight optimal features and their weights were selected (c). LASSO, least absolute shrinkage and selection operator Feature selection using 10-fold cross-validation LASSO. Eight optimal features and their weights were selected (c). LASSO, least absolute shrinkage and selection operator Table 5Detailed information of the 8 radiomic featuresCoefficientFeature full nameFilter typeFeature categoryPhysical meaning0.0594wavelet-HLL_firstorder_InterquartileRangeWavelet-HLLFirst-order Statistical FeatureReflects the dispersion and local heterogeneity of tumor gray distribution0.0436wavelet-LLH_glrlm_LongRunHighGrayLevelEmphasisWavelet-LLHSecond-order Texture Feature (GLRLM)Characterizes the local high-density texture of tumor lesions-0.0404wavelet-LHL_glcm_DifferenceEntropyWavelet-LHLSecond-order Texture Feature (GLCM)Reflects the chaos degree of intratumoral texture0.0358log-sigma-5-0-mm-3D_firstorder_90PercentileLog-sigma-5-0-mm-3DFirst-order Statistical FeatureReflects the upper limit of tumor gray intensity distribution0.0306log-sigma-5-0-mm-3D_firstorder_KurtosisLog-sigma-5-0-mm-3DFirst-order Statistical FeatureReflects the concentration of intratumoral gray intensity-0.0274wavelet-HHL_glcm_AutocorrelationWavelet-HHLSecond-order Texture Feature (GLCM)Reflects the local texture consistency of the tumor0.0217wavelet-HHH_glcm_ClusterShadeWavelet-HHHSecond-order Texture Feature (GLCM)Reflects the irregularity of intratumoral texture0.0008log-sigma-1-0-mm-3D_firstorder_RootMeanSquaredLog-sigma-1-0-mm-3DFirst-order Statistical FeatureCharacterizes the overall average pixel intensity of lesions Table 5Detailed information of the 8 radiomic featuresCoefficientFeature full nameFilter typeFeature categoryPhysical meaning0.0594wavelet-HLL_firstorder_InterquartileRangeWavelet-HLLFirst-order Statistical FeatureReflects the dispersion and local heterogeneity of tumor gray distribution0.0436wavelet-LLH_glrlm_LongRunHighGrayLevelEmphasisWavelet-LLHSecond-order Texture Feature (GLRLM)Characterizes the local high-density texture of tumor lesions-0.0404wavelet-LHL_glcm_DifferenceEntropyWavelet-LHLSecond-order Texture Feature (GLCM)Reflects the chaos degree of intratumoral texture0.0358log-sigma-5-0-mm-3D_firstorder_90PercentileLog-sigma-5-0-mm-3DFirst-order Statistical FeatureReflects the upper limit of tumor gray intensity distribution0.0306log-sigma-5-0-mm-3D_firstorder_KurtosisLog-sigma-5-0-mm-3DFirst-order Statistical FeatureReflects the concentration of intratumoral gray intensity-0.0274wavelet-HHL_glcm_AutocorrelationWavelet-HHLSecond-order Texture Feature (GLCM)Reflects the local texture consistency of the tumor0.0217wavelet-HHH_glcm_ClusterShadeWavelet-HHHSecond-order Texture Feature (GLCM)Reflects the irregularity of intratumoral texture0.0008log-sigma-1-0-mm-3D_firstorder_RootMeanSquaredLog-sigma-1-0-mm-3DFirst-order Statistical FeatureCharacterizes the overall average pixel intensity of lesions Detailed information of the 8 radiomic features Rad-score = 0.4298 + 0.0594×(wavelet-HLL_firstorder_Interquartile Range) + 0.0436×(wavelet-LLH_glrlm_LongRunHighGrayLevelEmphasis) + 0.0358×(log-sigma-5-0-mm-3D_firstorder_90Percentile) + 0.0306×(log-sigma-5-0-mm-3D_firstorder_Kurtosis) + 0.0217×(wavelet-HHH_glcm_ClusterShade) + 0.0008×(log-sigma-1-0-mm-3D_firstorder_RootMeanSquared) − 0.0274×(wavelet-HHL_glcm_Autocorrelation) − 0.0404×(wavelet-LHL_glcm_Difference Entropy) Three meaningful clinical features (age, CYFRA 21‑1, and EGFR-TKI grouping) were identified through univariate logistic regression (Table 6). The collinearity diagnosis showed that the variance inflation factors of all candidate variables were less than 5, with the maximum value being 1.028, indicating no interference from multicollinearity. Table 6Single factor logistic regression meaningful clinical characteristicsNumberVariableP-valueOR(95% CI)1Age0.0471.026 (1.000,1.053)2CYFRA 21‑10.0320.559 (0.328,0.952)3EGFR-TKIs*0.0061.426 (1.106,1.837)* EGFR‑TKIs were categorized as follows: first‑generation (gefitinib, erlotinib, icotinib); second‑generation (afatinib, dacomitinib); third‑generation (osimertinib, almonertinib, furmonertinib). CYFRA 21‑1, cytokeratin 19 fragment antigen 21‑1; EGFR-TKIs, epidermal growth factor receptor tyrosine kinase inhibitors; OR, odds ratio Table 6Single factor logistic regression meaningful clinical characteristicsNumberVariableP-valueOR(95% CI)1Age0.0471.026 (1.000,1.053)2CYFRA 21‑10.0320.559 (0.328,0.952)3EGFR-TKIs*0.0061.426 (1.106,1.837)* EGFR‑TKIs were categorized as follows: first‑generation (gefitinib, erlotinib, icotinib); second‑generation (afatinib, dacomitinib); third‑generation (osimertinib, almonertinib, furmonertinib). CYFRA 21‑1, cytokeratin 19 fragment antigen 21‑1; EGFR-TKIs, epidermal growth factor receptor tyrosine kinase inhibitors; OR, odds ratio Single factor logistic regression meaningful clinical characteristics * EGFR‑TKIs were categorized as follows: first‑generation (gefitinib, erlotinib, icotinib); second‑generation (afatinib, dacomitinib); third‑generation (osimertinib, almonertinib, furmonertinib). CYFRA 21‑1, cytokeratin 19 fragment antigen 21‑1; EGFR-TKIs, epidermal growth factor receptor tyrosine kinase inhibitors; OR, odds ratio The ROC curves of the clinical, radiomic, and combined models constructed based on RF in the training and test sets are presented in Fig. 6a, b. Figure 6c, d further clarifies the performance differences among the three models in predicting the therapeutic efficacy of EGFR-TKIs. The corresponding AUC, ACC, SEN, SPE, and F1 scores of the three predictive models in the training and test sets were calculated to assess their quantitative classification performance (Table 7). Fig. 6Performance comparison of the RF-based prediction models in the training and validation cohorts. ROC curve analysis and AUC (a, b) and decision curve analysis (c, d) in the training and validation sets. AUC, area under the curve; RF, random forest; ROC, receiver operating characteristic Performance comparison of the RF-based prediction models in the training and validation cohorts. ROC curve analysis and AUC (a, b) and decision curve analysis (c, d) in the training and validation sets. AUC, area under the curve; RF, random forest; ROC, receiver operating characteristic Table 7Performance comparison of the classification model in the training and test cohortsModelsTraining setTest setAUC(95% CI)ACCSPESENF1 scoreAUC(95% CI)ACCSPESENF1 scoreClinical0.720(0.647–0.783)0.6360.8460.3570.5920.668(0.517,0.806)0.6670.8440.4400.638Radiomics0.784(0.722,0.841)0.7190.7920.6220.7090.640(0.484,0.781)0.6320.7810.4400.608Fusion0.893(0.852,0.934)0.8070.8920.6940.7980.756(0.627,0.886)0.7020.8440.5200.683ACC, accuracy; AUC, area under the curve; SEN, sensitivity; SPE, specificity; CI, confidence intervals Table 7Performance comparison of the classification model in the training and test cohortsModelsTraining setTest setAUC(95% CI)ACCSPESENF1 scoreAUC(95% CI)ACCSPESENF1 scoreClinical0.720(0.647–0.783)0.6360.8460.3570.5920.668(0.517,0.806)0.6670.8440.4400.638Radiomics0.784(0.722,0.841)0.7190.7920.6220.7090.640(0.484,0.781)0.6320.7810.4400.608Fusion0.893(0.852,0.934)0.8070.8920.6940.7980.756(0.627,0.886)0.7020.8440.5200.683ACC, accuracy; AUC, area under the curve; SEN, sensitivity; SPE, specificity; CI, confidence intervals Performance comparison of the classification model in the training and test cohorts ACC, accuracy; AUC, area under the curve; SEN, sensitivity; SPE, specificity; CI, confidence intervals In the predictive model for EGFR-TKIs therapeutic efficacy within 1 year, the combined model based on CE-T1WI demonstrated the best performance compared to the radiomic and clinical models. The AUC performance of the combined model in the test set outperformed that of the radiomic and clinical models (0.756 vs. 0.644 vs. 0.668). The Delong test confirmed that the combined model achieved impressive predictive performance (Z-score: −2.166 vs. −4.640, P = 0.030 vs. < 0.001). It also obtained optimal metrics in the test set, with an AUC of 0.756 (95% CI: 0.627–0.886), ACC 0.702, SEN 0.844, SPE 0.520, and F1 score 0.683. In the external validation set, the ROC curve of the combined model is shown in Fig. 7a, with an AUC of 0.743 (95% CI: 0.604–0.881), ACC 0.679, SEN 0.480, SPE 0.857, and F1 score 0.662. These results confirmed the reliability of the combined model in predicting the therapeutic efficacy of EGFR-TKIs. The calibration curves of the combined model in the training, test, and external validation sets are shown in Fig. 7b, verifying the reliability and generalizability of the fused model. Fig. 7Performance of the fusion prediction model. ROC curve and AUC (a) in the external validation set, and calibration curve (b) in the training set, test set, and external validation set. AUC, area under the curve; ROC, receiver operating characteristic Performance of the fusion prediction model. ROC curve and AUC (a) in the external validation set, and calibration curve (b) in the training set, test set, and external validation set. AUC, area under the curve; ROC, receiver operating characteristic Discussion In this comprehensive study, we utilized multicohort data from patients with EGFR-mutant NSCLC BMs from three institutions to develop and externally validate an MRI radiomic framework for predicting the therapeutic response to EGFR-TKIs. Intratumoral radiomic features extracted from CE-T1WI show favorable predictive performance for EGFR‑TKIs response, and our exploratory analysis of 3/5/8 mm peritumoral regions provides a useful reference for future related studies. We discovered a synergistic effect between radiomic features and clinical characteristics. The integration of clinical features and MRI radiomics ultimately facilitated the development of a combined model that used multiple data sources to optimize predictive performance. The validation framework underscored the pivotal role of radiomic features extracted from MRI. These features provide complementary insights to established MRI biomarkers, bringing us closer to achieving personalized treatment strategies for NSCLC BMs. Compared to traditional visual assessment methods based on enhanced MRI of BMs, the radiomic model exhibited significant advantages in terms of accuracy and sensitivity in predicting the therapeutic response of BMs, thereby aiding clinicians in formulating more personalized treatment plans and promoting the application and development of radiomic approaches. We constructed and externally validated a combined clinical-radiomic predictive model, which demonstrates promising potential for assisting treatment efficacy prediction and clinical decision-making. This model may provide clinicians with objective and referential evidence to facilitate the formulation of individualized treatment strategies and follow-up plans. Further prospective validation is still required before its routine clinical application. The motivation for using multiple classifiers was to enhance the robustness of the results, as previous studies have indicated that selecting an appropriate algorithm plays a crucial role in classification outcomes [20, 21]. Zhao et al. [22] constructed support vector machine, RF, and k-nearest neighbors models using LASSO based on CE-T1WI and T2-fluid-attenuated inversion recovery sequences to differentiate between primary central nervous system lymphoma and BMs. The RF model exhibited the best performance (AUC = 0.73). Kanakarajan et al. [23] developed a radiomic model based on CE-T1WI to predict the efficacy of radiotherapy for BMs. By incorporating clinical features, the combined model using an RF classifier achieved the highest AUC of 0.89, with an accuracy rate of 87%. Similar studies [24] have shown that radiomic models constructed using an RF classifier also achieved favorable performance in predicting the efficacy of radiotherapy for BMs, with an AUC of up to 0.83. In summary, the RF classifier has demonstrated excellent performance in predicting the therapeutic efficacy for BMs, which agrees with the findings of our study. Our model training also indicated the significant efficacy of the RF classifier. Previous studies have shown that extracting radiomic features from the tumor region and peritumoral edema of BMs is reasonable and reliable. This approach can encompass most of the details in the heterogeneous regions of BMs and provide better prediction of therapeutic efficacy in patients with BMs, compared with features from a single region [25]. Fan et al. [26] also indicated that the tumor region and peritumoral edema in NSCLC BMs were complementary, allowing for the extraction of rich radiomic phenotypic features to enhance model robustness and achieve more reliable prediction of the therapeutic efficacy of EGFR-TKIs. Based on previous studies, we constructed models using intratumoral and peritumoral features based on the optimal ML algorithm. However, the results indicated that the model constructed using features extracted from within the tumor exhibited better performance. This may have been because NSCLC patients with EGFR mutations often develop multiple small BMs, with mild or no peritumoral edema [27]. Therefore, it is impossible to extract critical information from the peritumoral edema area pertaining to the therapeutic efficacy against BMs. Among patients with NSCLC BMs, those with EGFR mutations have longer survival and better control of intracranial disease [28]. Fan et al. [26] incorporated pretreatment CE-T1WI and T2WI from NSCLC BMs patients and constructed models based on tumor area analysis, peritumoral edema, margin/background parenchyma, and multiregional fusion to predict the therapeutic response to EGFR-TKIs. The fusion model achieved the best performance in the external validation set, proving that subregional radiomics, as a novel noninvasive approach, can help guide personalized treatment strategies with EGFR-TKIs for NSCLC BMs. A predictive model combining radiomics with clinical predictive features will enable earlier detection of disease progression and benefit patients by guiding clinicians to conduct more intensive MRI follow-up of patients at high risk of intracranial progression. Similarly, Qu et al. [29] included 212 lesions from 70 patients with NSCLC BMs who received first-line EGFR-TKIs treatment. Radiomic features were extracted from the brain tumor regions on pretreatment CE-T1WI, and a radiomic score was calculated based on the selected features. The combined predictive model incorporating EGFR-19del mutation, third-generation EGFR-TKI treatment, and the average rad-score outperformed models constructed using each of these three predictors alone. The combined model demonstrated good predictive value for intracranial progression within 1 year after EGFR-TKIs treatment. The results confirmed that the nomogram combining clinical and radiomic features outperformed nomograms based solely on either features alone. Unlike the aforementioned validation studies, in our study, age, EGFR-TKIs, and CYFRA 21‑1 were identified as significant clinical features. CYFRA 21‑1 was the most sensitive tumor marker in NSCLC, which is consistent with the literature [30]. Qi et al. [31] included 117 patients with NSCLC BMs for individualized prediction of the therapeutic efficacy of first-generation EGFR-TKIs. The LASSO algorithm was used to screen radiomic features extracted from multiparametric MRI. In the short-term efficacy predictive model, clinical, radiomic, and combined nomograms were constructed separately. The results demonstrated good consistency between predicted risk and actual outcomes, confirming that the combined nomogram outperformed nomograms based solely on clinical or radiomic features, with a C-index of 0.843. Qi et al. [32] subsequently developed and validated a nomogram based on clinical characteristics and MRI radiomics for predicting the short-term efficacy of third-generation EGFR-TKIs in patients with EGFR-mutant lung adenocarcinoma BMs, with a C-index of 0.803. The studies confirmed the favorable efficacy of first- and third-generation EGFR-TKIs in treating EGFR-mutant NSCLC BMs. The predictive performance of nomograms established by combining radiomics and clinical features outperformed nomograms based solely on radiomics or clinical features, and served as a noninvasive predictive tool that aided timely personalized adjustment of treatment plans. Association of radiomic characteristics with clinical features is an emerging field that enhances the correlation between radiomic features and molecular characteristics, potentially reflecting gene expression differences within and between tumors. The aforementioned studies indicate that MRI radiomics combined with clinical features can provide noninvasive and intelligent prediction of the therapeutic efficacy of EGFR-TKIs in EGFR-mutant NSCLC BMs. However, because of the design of thecase follow-up, the research samples were from a single center and limited in number, which restricts the generalizability and clinical effectiveness of the results. Despite the use of cross-validation and the good performance of the models in the training and test sets, the generalizability of the models to unseen data requires further investigation. In this study, radiomic analysis was performed using pretreatment CE-T1WI from NSCLC BM patients. To maximize the sample size, patients with different EGFR mutation types receiving different EGFR-TKIs were included. Based on a multi-institutional and large sample size, combined with clinical data, the effectiveness of EGFR-TKIs in local tumor control was predicted. The results indicated that radiomic features combined with clinical features served as complementary biomarkers for predicting the therapeutic efficacy in NSCLC BM patients, achieving an AUC of 0.743 in the external validation set, demonstrating good generalizability. Our study demonstrated that combining MRI radiomics with clinical features improved the accuracy of EGFR-TKIs therapeutic monitoring based on imaging, and supported preoperative diagnosis, treatment planning, and outcome determination of BMs. Our study has several noteworthy limitations that should be acknowledged when interpreting the results. First, despite being a multicenter study with an independent external validation cohort, our predictive model was constructed based on retrospective clinical and imaging data. Such a retrospective design inevitably carries inherent selection biases and makes it difficult to fully control for potential confounding factors; thus further prospective investigations are required to validate the model and refine its clinical applicability. Second, EGFR mutation genotyping was performed using primary lung cancer tissue samples, which may not fully represent the secondary genetic alterations that occur in intracranial metastatic brain lesions. Third, there existed heterogeneity in MRI scanning devices and imaging parameters across included centers; meanwhile, enrolled patients received heterogeneous treatment regimens and were at different clinical stages of disease, all of which might weaken the generalizability of the study conclusions. Fourth, the radiomic analysis was solely based on a single contrast-enhanced T1-weighted imaging (CE-T1WI) sequence, as some eligible patients lacked other standard brain magnetic resonance sequences, including diffusion-weighted imaging (DWI) and diffusion tensor imaging (DTI), which are routinely utilized for BM assessment. Expanding the inclusion of multimodal MRI sequences in future research may help further optimize the predictive performance of the model. Finally, the current model calibration still needs to be improved, and future work is warranted to integrate multi-modal imaging and multi-dimensional clinical data for methodological optimization. Moreover, in-depth mechanistic analyses focusing on the correlation between radiomic features and tumor biological behaviors will help clarify the underlying pathogenesis of BM and treatment response, ultimately promoting the clinical translation of the proposed predictive model. Conclusion This study developed and externally validated a combined clinical-radiomic model for predicting EGFR-TKI therapeutic efficacy in patients with NSCLC brain metastases. The model demonstrated acceptable predictive performance in external validation, suggesting potential value for supporting individualized treatment evaluation and clinical decision-making. Nevertheless, the predictive performance of the model remains moderate, and its generalizability needs to be further verified and improved in future prospective studies with larger cohorts.