Interpretable habitat radiomics based nomogram for predicting T790M mutation in non-small cell lung cancer with brain metastases

Abstract

Background

Spatial heterogeneity within the tumor drives T790M resistance mutation. This study aimed to develop a habitat radiomics-based nomogram for predicting T790M resistance mutation for non-small cell lung cancer (NSCLC) patients with brain metastases (BM).

Methods

A total of 301 patients from Beijing Chest Hospital, Capital Medical University and Shandong Cancer Hospital and Institute, Shandong First Medical University were retrospectively collected. Tumors were manually delineated on contrast-enhanced T1-weighted (T1-CE), and the segmentation was combined with an automatically generated 5 mm peritumoral region to form the whole region of interest (ROI). Predictive models were established using radiomics features extracted from intratumoral, peritumoral, whole ROI, and habitat-generated subregions within the whole ROI. A nomogram was developed by integrating a representative radiomics signature and meaningful clinical factors. Model performance was assessed using the area under the receiver operating characteristic (ROC) curve (AUC) and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) was used to interpret feature distribution in the habitat model, and proteomics analysis was performed to explore the biological mechanisms of the nomogram.

Results

The habitat model based on three subregions achieved higher AUCs of 0.924 and 0.902 in the training and test cohorts, respectively. The habitat radiomics-based nomogram yielded the highest AUCs of 0.943 and 0.931 in the two cohorts, separately. Meanwhile, clinical and other radiomics models showed moderate performance with AUCs ranging from 0.644 to 0.800 in the test cohort. Proteomics analysis revealed that RNA processing and splicing may represent the most critical pathway distinguishing high- and low-risk groups.

Conclusions

Habitat imaging combining tumoral and peritumoral radiomics was beneficial for identifying T790M resistance mutation in NSCLC BM patients. Furthermore, a habitat radiomics-based nomogram can further improve the predictive ability.

Introduction

Lung cancer remains the most prevalent malignant disease, causing more deaths worldwide than any other cancer (1). Non-small cell lung cancer (NSCLC) accounts for nearly 85% of lung cancer, and patients harboring actionable oncogenic mutations are candidates for targeted therapy (2). Epidermal growth factor receptor (EGFR) mutation is the most common genetic alteration detected in NSCLC (1). Although EGFR tyrosine kinase inhibitors (TKIs) have significantly improved progression-free survival, the acquisition of the T790M resistance mutation represents an inevitable challenge during treatment with first- or second-generation TKIs (3). The third-generation inhibitor osimertinib demonstrates specific activity against this resistant isoform and is thus recommended for the treatment of T790M-mutated tumors (4). The development of metastasis is frequently associated with acquired T790M resistance mutation (3). Up to 60% of NSCLC patients will develop brain metastases (BM) during disease progression (5). As the primary tumor may have been resected or otherwise treated, molecular analysis of metastatic sites can offer a more definitive means for identifying underlying resistance genes.

Re-biopsy of the primary tumor or metastatic lesions is considered the gold standard for identifying T790M resistance mutation, yet this approach is invasive and may not be suitable for the patient’s physical condition (6). Liquid biopsy is a common alternative for detecting T790M mutation in clinical practice. However, it has significant limitations, particularly when low disease burden affects the accuracy of results (3). Magnetic resonance imaging (MRI) is a routine modality for monitoring treatment response in BM patients, but it is challenging for radiologists to visually assess mutation status based solely on conventional MRI. Recently, radiomics has emerged as a promising tool that can provide non-invasive biomarkers reflecting tumor molecular characteristics by quantifying intratumoral heterogeneity (7). Radiomics has been widely applied in various aspects of BM management, including diagnosis, treatment response assessment, and prognostic prediction (8-10). Furthermore, an increasing number of studies have confirmed that features derived from the peritumoral region can also provide complementary information for further improving diagnostic performance (11,12).

Tumor growth may exhibit regional phenotypic variability, influenced by factors such as tumor location, interactions with surrounding blood vessels, and host immune response (13,14). Given this intratumoral heterogeneity, the concept of habitat imaging has been proposed. Habitat imaging enables the qualitative and quantitative characterization of spatially heterogeneous subregions within tumors, which may present distinct treatment responses and regional differences in genotype and phenotype (15,16). Radiomics features derived from habitat imaging have demonstrated advantages in the analysis of various tumors, including glioblastoma, breast cancer, and lung cancer (17,18).

In this study, we first aimed to construct and validate a habitat model combining intratumoral and peritumoral radiomics features for identifying the presence of T790M mutation, and to quantify features contributions to the model using SHapley Additive exPlanation (SHAP). Second, by integrating clinical predictive factors, we aimed to establish a habitat radiomics-based nomogram for accurate prediction, and further explore the underlying biological mechanism using proteomics analysis. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0244/rc).

Methods

Participants

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The ethical approvals were obtained from the Institutional Review Boards of Beijing Chest Hospital, Capital Medical University and Shandong Cancer Hospital and Institute, Shandong First Medical University. The need for patient informed consent was waived due to the retrospective nature of this study.

Patients diagnosed with NSCLC were enrolled based on the following criteria: (I) clinically or histopathologically confirmed BM; (II) confirmed EGFR mutation status; (III) availability of T790M mutation test results; and (IV) pretreatment high-resolution MRI scans obtained prior to any therapeutic intervention. The exclusion criteria were as follows: (I) incomplete clinical data; (II) prior malignancy history; and (III) coexisting other central nervous system diseases. A total of 301 patients from two centers were ultimately enrolled in accordance with the aforementioned inclusion and exclusion criteria. Between January 2015 and June 2021, 211 NSCLC patients with BM were recruited from Beijing Chest Hospital, Capital Medical University, including 142 patients without T790M mutation and 69 patients with T790M mutation. Additionally, 60 patients without T790M mutation and 30 patients harboring T790M mutation were recruited from Shandong Cancer Hospital and Institute, Shandong First Medical University between January 2020 and January 2021. The enrollment process is summarized in Figure 1.

Figure 1.

The enrollment process of this study. Center 1: Beijing Chest Hospital, Capital Medical University; Center 2: Shandong Cancer Hospital and Institute, Shandong First Medical University. BM, brain metastases; NSCLC, non-small cell lung cancer.

MRI protocol image preprocessing

Pre-treatment MRI examinations were performed at the two centers using one of four available MRI scanners. The sequences mainly included contrast-enhanced T1-weighted (T1-CE) fast spin-echo imaging. Intravenous gadolinium-based contrast agent was used at a dose of 0.1 mmol/kg. Detailed scanning parameters for T1-CE are provided in Table S1.

N4 bias field correction was applied to the T1-CE sequence to reduce bias field artifacts. The T1-CE images were then resampled to 1×1×1 mm³ isotropic voxels to standardize voxel spacing.

BM segmentation

Segmentation was conducted by a radiologist with at least 5 years of experience in brain MRI and cross-validated by an independent radiologist with at least 6 years of experience. For patients with multiple tumors, the largest lesion was manually contoured and used as the region of interest (ROI) for subsequent radiomics analysis, defined as the intratumoral region of interest (ROI_I). Both radiologists were blinded to the gene mutation status of BM. Discrepancies between the two independent delineations were resolved via joint discussion and consensus adjudication. The final consensus ROI_I mask was adopted for subsequent radiomics feature extraction, rather than selecting either individual observer segmentation or simply averaging the two masks. Segmentation was performed using 3D Slicer (http://www.slicer.org). To quantify inter-rater reliability, the Dice similarity coefficient (DSC) and intraclass correlation coefficient (ICC) were calculated for the delineated ROI_I between the two radiologists. The average DSC was 0.89 (range, 0.82–0.94) and the ICC was 0.92 [95% confidence interval (CI): 0.88–0.95], indicating excellent inter-observer agreement and confirming the reliability of the manual segmentation.

Peritumoral region generation

The ROI was dilated outward using elliptical kernels to enlarge the tumor mask based on pixel spacing to augment the areas and dimensions of BM regions. The peritumoral region of interest (ROI_P) was explored by configuring dilation intervals of 5 mm to assess the predictive power. Then, the intratumoral and peritumoral regions were combined as the whole ROI.

Subregions of habitat generation in whole ROI

The habitat refers to functionally distinct subregions within the whole ROI, which are generated based on the similarity of voxel-wise radiomics features and reflect intratumoral heterogeneity by capturing spatial differences in imaging characteristics. Radiomics features were characterized for each voxel by combining both local and global imaging information. This extraction captures a diverse set of 19 radiomics features, including entropy, median, difference average, and other features, as shown in Figure 2. The selection of these specific features followed Image Biomarker Standardization Initiative (IBSI) standard compliance, biological and imaging pertinence, and low redundancy and high robustness. Then, stringent quality control was performed prior to clustering analysis, which included normalization, standardization, and retention of features with ICC >0.75 indicating good reliability.

Figure 2.

The generated habitat regions and 19 characteristics are present. ROI, region of interest.

Pixels with identical cluster labels exhibit similarities in intensity and neighborhood texture characteristics, where the dispersion of clustering patterns in the label map reflected the degree of intratumoral heterogeneity. Locally, entropy features were computed for all voxels within the mask region to quantify spatial complexity. Three tumor habitats were generated via unsupervised clustering of the aforementioned voxel-wise radiomics features within the whole ROI. Specifically, the K-means method was used to partition tumor voxels into K optimal spatial subregions across the image. The number of cluster centers was pre-specified to range from 3 to 10 to categorize potential unique habitat regions within the whole ROI. The optimal number of clusters was selected using three evaluation metrics: Calinski-Harabasz score, Silhouette score, and Davies–Bouldin score. These metrics ensure the rationality of habitat partitioning by evaluating the compactness within clusters and the separation between clusters.

Radiomics feature extraction

Prior to feature extraction, all imaging datasets were subjected to z-score intensity normalization and gray-level discretization. Specifically, z-score normalization was implemented on a whole-dataset cohort level rather than per patient. Gray-level discretization was then performed using fixed bin widths to standardize intensity distributions among individuals. Radiomics features were computed from the original images and seven filter-transformed images. A comprehensive feature set, including first-order statistics (intensity distribution metrics), shape descriptors (3D geometric properties), and texture features (e.g., GLCM, GLRLM), was extracted from six regions (ROI_I, ROI_P, whole ROI, habitat 1, habitat 2, and habitat 3). In total, 1,834 radiomics features were generated per region, with detailed mathematical definitions and implementation parameters complying with the IBSI guidelines, as documented in the PyRadiomics reference (v3.0.1).

Feature selection

The synthetic minority oversampling technique (SMOTE) was first applied to balance the number of patients with and without T790M resistance mutation in the training cohort, avoiding potential model bias towards the majority class. To retain only robust features for modeling, radiomics features were first filtered based on inter-rater reliability: features with ICC >0.75 were retained, whereas those with ICC ≤0.75 were excluded. All subsequent feature screening steps were nested within 5-fold cross-validation (CV). Specifically, in each iteration of 5-fold CV, the dataset was split into a training and a test cohort (one fold), and all feature selection procedures were conducted exclusively on the training set without accessing test cohort data. Firstly, the independent samples t-test (two-tailed) was employed to assess between-group differences in radiomics features. Features with P<0.05 were retained after Bonferroni correction to reduce type I error inflation caused by multiple comparisons. Secondly, Pearson’s correlation coefficient was calculated pairwise for the retained features to identify linear dependencies. For features with |r|>0.9, the one with stronger discriminative power was retained to eliminate collinearity. Thirdly, the least absolute shrinkage and selection operator (LASSO) model with L1-norm penalty was implemented to select the most predictive features with non-zero coefficients. These selected features were then validated on the corresponding test cohort within each cross-validation fold. Finally, the rad-score was calculated as a linear weighted combination of the selected features and their corresponding regression coefficients.

Model construction

Based on the representative radiomics signature selected via the nested CV-based feature selection strategy, we developed radiomics models through comprehensive machine learning algorithms. Specifically, seven distinct regional models were constructed for ROI_I, ROI_P, whole ROI, habitat 1, habitat 2, habitat 3, and the clinical model, respectively, followed by an integrated model: a combined habitat model incorporating habitat 1, 2, and 3. To establish these predictive models, we implemented seven machine learning algorithms: Logistic Regression, Support Vector Machine, Random Forest, Extremely Randomized Trees, eXtreme Gradient Boosting, Light Gradient Boosting Machine, and Multilayer Perceptron, systematically exploring their predictive capabilities across different spatial configurations of tumor subregions. Finally, we constructed a nomogram by incorporating the rad-score of the habitat model and important clinical factors. The prediction nomogram was established based on multivariate logistic regression. First, the rad-score was calculated as a linear combination of selected key radiomics features weighted by their corresponding regression coefficients. Next, independent clinical risk factors identified via multivariate analysis and the rad-score were integrated into the final combined regression model. In the process of model construction, the overall cohort was strictly split into a training and an independent test cohort based on two different centers. The test cohort was completely locked and isolated in advance. All candidate models and hyperparameters were optimized and screened solely using the training cohort. The optimal model was determined based on the performance of the test cohort. The workflow pipeline is shown in Figure 3.

Figure 3.

The workflow pipeline of this study. DCA, decision curve analysis; FC, fold change; ROC, receiver operating characteristic; ROI, region of interest; SHAP, SHapley Additive exPlanation.

Interpretability analysis

SHAP was used to assess feature contribution of the habitat model and enhance its interpretability. Moreover, 20 patients from the test cohort underwent surgical resection after being diagnosed with BM. To further elucidate the intrinsic biological mechanisms underlying the ability of nomogram to distinguish T790M resistance mutation, this study collected formalin-fixed paraffin-embedded samples from these patients. Based on the cut-off value of the nomogram, these patients were stratified into high- and low-risk groups and subjected to proteomic analysis to identify the differentially expressed proteins (DEPs). Detailed analytical procedures are provided in the Appendix 1.

Statistical analysis

Statistical analyses were conducted using the software SPSS 26.0 (IBM Corp., Armonk, NY, USA) and Python’s Scikit-learn package. Independent samples t-test was employed for continuous variables (two-group comparisons), and χ² test was used for categorical variables to assess intergroup differences in clinical characteristics. Model performance was evaluated using receiver operating characteristic (ROC) curve analysis with area under the curve (AUC) quantification (95% CI), complemented by classification metrics including accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Decision curve analysis (DCA) was used to evaluate the clinical utility of these models. A P<0.05 was considered statistically significant throughout the entire analytical framework.

Results

Clinical characteristics of patients

The clinical characteristics of all patients are summarized in Table 1. As shown in Table 1, EGFR mutation and TKI treatment duration have significant differences between patients with and without T790M mutation, whether in the training or test cohorts. In contrast, age, gender, and smoking history showed no significant differences between the two groups both in the training and test cohorts.

Table 1. The clinical characteristics of NSCLC BM patients in the training and test cohorts.

Characteristics	Training cohort (n=211)				Test cohort (n=90)
	EGFR (n=142)	EGFR T790M (n=69)	P value		EGFR (n=60)	EGFR T790M (n=30)	P value
Age (years)	64.02±9.08	61.94±11.76	0.159		62.67±8.67	59.50±15.10	0.209
Gender			0.676				0.179
Male	78 (54.93)	40 (57.97)			31 (51.67)	11 (36.67)
Female	64 (45.07)	29 (42.03)			29 (48.33)	19 (63.33)
Smoking			0.171				0.058
Yes	66 (46.48)	39 (56.52)			16 (26.67)	14 (46.67)
No	76 (53.52)	30 (43.48)			44 (73.33)	16 (53.33)
EGFR mutation			<0.001*				0.003*
Exon 19 deletion	45 (31.69)	58 (84.06)			18 (30.00)	20 (66.67)
Exon 21 L858R mutation	93 (65.49)	10 (14.49)			38 (63.33)	10 (33.33)
Others	4 (2.82)	1 (1.45)			4 (6.67)	0
TKI treatment duration			<0.001*				<0.001*
≤12 months	99 (69.72)	24 (34.78)			43 (71.67)	9 (30.00)
>12 months	43 (30.28)	45 (65.22)			17 (28.33)	21 (70.00)

Data are presented as n (%) or mean ± standard deviation. *, P<0.05. BM, brain metastases; EGFR, epidermal growth factor receptor; NSCLC, non-small cell lung cancer; TKI, tyrosine kinase inhibitor.

Feature selection

Finally, the optimal number of clusters was determined to be 3 based on the Calinski-Harabasz score, Silhouette score, and Davies-Bouldin score. Table S2 presents these scores for different number of cluster centers. Habitat 1 was primarily distributed within the enhancing solid tumor core. Both habitat 2 and habitat 3 occupied the peritumoral region. Habitat 3 referred to the peritumoral region adjacent to the tumor, whereas habitat 2 represented the distal peritumoral region. From ROI_I, ROI_P, and the whole ROI, 23, 20, and 32 radiomics features were extracted, respectively. Based on the subregions of habitat 1, habitat 2, and habitat 3, 28, 21, and 26 features were extracted for model construction, respectively. Besides, the habitat model included 9, 8 and 11 features from habitat 1, habitat 2, and habitat 3, respectively. The coefficients of features in different models are shown in Figure S1.

Model performance

To establish robust predictive models, we adopted the most optimal machine learning algorithm, with the highest AUC in the test cohort as the primary selection criterion. Region-specific model performance revealed distinct algorithmic preference: extremely randomized trees achieved optimal performance for habitat 1, habitat 3, ROI_I, whole ROI, and the habitat model; multilayer perceptron was optimal for habitat 2; support vector machine performed best for ROI_P and also emerged as the superior algorithm for the clinical model. The nomogram was constructed by integrating EGFR mutation, TKI treatment duration, and the rad-score derived from the habitat model (Figure 4). This model achieved the best predictive performance, with peak AUC values of 0.943 (95% CI: 0.918–0.967) and 0.931 (95% CI: 0.883–0.980) in the training and test cohorts, respectively. The habitat model maintained strong discriminative capability, with AUCs of 0.924 (95% CI: 0.895–0.953) and 0.902 (95% CI: 0.841–0.963) in the two cohorts. At the clinically relevant threshold, the nomogram achieved PPV of 0.855 and NPV of 0.871 in the training cohort, and 0.651 and 0.957 in the test cohort, respectively. For the habitat model, the corresponding PPV and NPV were 0.863 and 0.810 in the training cohort, and 0.700 and 0.960 in the test cohort. Regional model evaluations showed consistent performance trends: habitat 1, habitat 2, habitat 3, ROI_I, ROI_P, whole ROI, and clinical models achieved training AUCs of 0.893 (95% CI: 0.855–0.930), 0.803 (95% CI: 0.752–0.854), 0.884 (95% CI: 0.846–0.921), 0.894 (95% CI: 0.859–0.929), 0.864 (95% CI: 0.820–0.908), and 0.906 (95% CI: 0.873–0.940), which were subsequently validated with AUCs of 0.873 (95% CI: 0.802–0.944), 0.719 (95% CI: 0.615–0.824), 0.817 (95% CI: 0.730–0.904), 0.861 (95% CI: 0.782–0.939), 0.758 (95% CI: 0.656–0.859), and 0.883 (95% CI: 0.815–0.952), respectively, in the test cohort. The clinical model achieved AUCs of 0.714 (95% CI: 0.654–0.775) and 0.676 (95% CI: 0.567–0.785) in the training and test cohort, respectively. The performance of all models is shown in Table 2. Comparative ROC and DCA curves for all models are presented in Figure 5 and Figures S2-S4. The calibration curves of clinical model, habitat model, and nomogram are presented in Figure S5.

Figure 4.

The habitat radiomics based nomogram to predict the T790M resistance mutation.

Table 2. The performance of different models in predicting T790M.

Cohort	Models	AUC (95% CI)	Accuracy	Sensitivity	Specificity
Training	Intratumoral	0.894 (0.859, 0.929)	0.810	0.980	0.711
	Peritumoral	0.864 (0.820, 0.908)	0.810	0.739	0.880
	Whole	0.906 (0.873, 0.940)	0.824	0.775	0.873
	Habitat 1	0.893 (0.855, 0.930)	0.827	0.754	0.901
	Habitat 2	0.803 (0.752, 0.854)	0.750	0.838	0.662
	Habitat 3	0.884 (0.846, 0.921)	0.796	0.732	0.859
	Habitat	0.924 (0.895, 0.953)	0.835	0.796	0.873
	Clinical	0.714 (0.654, 0.775)	0.690	0.697	0.683
	Nomogram	0.943 (0.918, 0.967)	0.863	0.873	0852
Test	Intratumoral	0.861 (0.782, 0.939)	0.744	0.833	0.700
	Peritumoral	0.758 (0.656, 0.859)	0.644	0.900	0.517
	Whole	0.883 (0.815, 0.952)	0.800	0.867	0.767
	Habitat 1	0.873 (0.802, 0.944)	0.789	0.867	0.750
	Habitat 2	0.719 (0.615, 0.824)	0.644	0.767	0.853
	Habitat 3	0.817 (0.730, 0.904)	0.756	0.733	0.767
	Habitat	0.902 (0.841, 0.963)	0.844	0.933	0.800
	Clinical	0.676 (0.567, 0.785)	0.689	0.667	0.700
	Nomogram	0.931 (0.883, 0.980)	0.811	0.933	0.750

AUC, area under the curve; CI, confidence interval.

Figure 5.

The ROC curves of different models in the training and test cohorts. (A-C) ROC curves of the intratumoral, peritumoral, whole, habitat 1, habitat 2, habitat 3, habitat, clinical, and nomogram models in the training cohort, respectively. (D-F) ROC curves of these models in the test cohort, respectively. ROC, receiver operating characteristic.

Interpretability analysis

The contributions of radiomics features to the positive and negative predictions of patients are displayed in a summary plot in Figure 6A. It was found that lbp_3D_ml_ngtdm_Contrast_h3 is the feature with the greatest contribution in the habitat model. The cut-off value for the nomogram was set at 0.542. Based on this threshold, proteomic analysis identified 382 DEPs between the high- and low-risk groups (Figure 6B). In this exploratory proteomic analysis, Gene Ontology (GO) enrichment analysis revealed that the DEPs are predominantly involved in RNA processing, nuclear protein-containing complex, nucleic acid binding, RNA metabolic process, nucleic acid metabolic process, and RNA binding (Figure 6C). Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis showed potential associations of the DEPs with the spliceosome, ATP-dependent chromatin remodeling, RNA polymerase, and nucleotide excision repair (Figure 6D). Gene Set Enrichment Analysis (GSEA) based on the GO database provided preliminary evidence that the high-risk group is mainly characterized by up-regulated pathways. The top five pathways ranked by enrichment score are active ion transmembrane transporter activity, ATPase coupled ion transmembrane transporter activity, cation-transporting ATPase activity, proton-transporting two-sector ATPase complex catalytic domain, ATPase activity coupled to transmembrane movement of ions, rotational mechanism (Figure S6).

Figure 6.

Interpretability of the model. (A) SHAP summary plots of the habitat model. (B) Volcanic plot shows the DEPs between the high-risk and low-risk groups based on the nomogram. (C) Bubble plot of GO enrichment analysis. (D) Bubble plot of KEGG enrichment analysis. DEPs, differentially expressed proteins; FC, fold change; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; SHAP, SHapley Additive exPlanation.

Discussion

Intratumoral heterogeneity is a distinctive characteristic of malignant tumors, contributing to differences in tumor growth rate, invasion, metastasis, and prognosis (19). It can also drive tumor adaptation and resistance during treatment, leading to rapid disease progression, treatment failure, and ultimately reduced survival rates in cancer patients (20). A comprehensive understanding of the causes and phenotypic variations in tumor progression is essential for achieving personalized treatment strategies. T790M resistance mutation is an important manifestation of heterogeneity in NSCLC BM patients. Timely identification of the presence of T790M resistance mutation in clinical practice can facilitate the adjustment of treatment strategies. Habitat imaging can segment the tumor microenvironment into distinct subregions with similar characteristics, thereby providing insights into the distribution of intratumoral heterogeneity (19). Additionally, given the aggressive nature of malignant tumors and their ability to invade surrounding brain tissue, the peritumoral region may also have potential value for evaluating genetic status. A 5 mm peritumoral region has been widely validated to capture peritumoral microenvironmental changes associated with tumor invasion, edema, and molecular status (21-23). Therefore, this study developed a novel approach integrating habitat imaging and peritumoral radiomics to build a habitat model for identifying NSCLC BM patients harboring the T790M resistance mutation.

In our study, patients with T790M mutation were younger than those without the resistance mutation in both the training and test cohorts. Previous studies have demonstrated that patients younger than 65 years of age are significantly associated with the acquisition of the T790M mutation at the time of disease progression (24,25). Our results also indicated that patients with a longer treatment duration were more likely to develop T790M resistance mutation. Lin et al. revealed that an EGFR TKI treatment duration of more than 13 months is an independent factor for the development of secondary T790M mutation (26). In our cohort, differences remained between NSCLC BM patients harboring different subtypes of EGFR mutation. Patients with EGFR exon 19 deletion may be more likely to develop T790M mutation compared to those with EGFR L858R mutation, whereas those with uncommon EGFR mutation were less likely to develop resistance mutation. Prior studies have also shown that the incidence of T790M is higher in patients with EGFR exon 19 deletion than in those with EGFR L858R mutation, and significantly higher than in patients without sensitive EGFR mutation (26,27). Based on baseline mutation subtypes and treatment duration, the clinical model exhibited moderate performance, with AUCs of 0.786 and 0.783 in the training and test cohorts, respectively.

Recently, radiomics has demonstrated potential value in predicting mutation status in NSCLC patients with BM (28-31). Wang et al. constructed an MRI-based radiomics signature to predict EGFR mutation status in lung adenocarcinoma with BM (31). They found that the radiomics features derived from T2 fluid-attenuated inversion recovery sequences were strongly correlated with EGFR mutation status. Previous studies have also demonstrated that radiomics features can also effectively predict T790M resistance mutation (Table 3) (32-35). Li et al. developed and validated radiomics models using multi-sequence MRI to identify the emergence of the T790M mutation (34). Their study indicated that the model built by diffusion weighted imaging achieved the best performance, with AUCs of 0.886 and 0.850 in the training and test cohorts, respectively. Fan et al. delineated the tumor active area and peritumoral edema area, and subsequently built radiomics signature based on these two regions to assess the presence of T790M mutation following treatment (35). Their study revealed that a multi-region combined radiomics model could effectively detect the T790M mutation. Habitat imaging has been shown to offer rich information about tumor behavior and is widely used in various cancers (36-40). Pan et al. reported that a subregion-based model could effectively predict the risk of local recurrence and support precision therapy in head and neck squamous cell carcinoma (39). Shang et al. (40) developed a habitat model integrating radiomics features derived from both tumor and peritumoral region to identify the invasiveness of lung adenocarcinoma. However, the application of habitat imaging in NSCLC patients with BM remains limited. Compared with previous studies using whole ROI, dual‑region, or brain‑tumor interface radiomics, we are the first to develop a habitat radiomics based model by unsupervised clustering of intratumoral and peritumoral subregions, enabling more comprehensive characterization of spatial heterogeneity.

Table 3. The summary of the relevant studies.

Study	Modality	Sequences	ROI	AUC (training/test)	Biological validation
Wu, 2024 (32)	1.5T MRI	T1-CE, T2WI, T2-FLAIR, DWI	Whole tumor	0.89/N/A	None
Fan, 2023 (33)	3.0T MRI	T1-CE, T2WI	Brain‑tumor interface + peritumoral edema	0.818/0.769	None
Li, 2022 (34)	3.0T/1.5T MRI	T1-CE, T2WI, T2-FLAIR, DWI	Whole tumor	0.886/0.850	None
Fan, 2022 (35)	3.0T MRI	T1-CE, T2WI	Tumor active area + peritumoral edema	0.938/0.860	None

AUC, area under the curve; DWI, diffusion-weighted imaging; MRI, magnetic resonance imaging; N/A, not applicable; T1-CE, contrast-enhanced T1-weighted; T2WI, T2-weighted imaging; T2-FLAIR, T2 fluid-attenuated inversion recovery.

In this study, we extracted radiomics features from the tumor and peritumoral 5 mm region and then constructed radiomics models based on these features, respectively. It found that the intratumoral radiomics performed better than the peritumoral radiomics with the AUC of 0.861 in the test cohort. Further, a whole model was built by integrating tumor and peritumoral region. The performance of the whole model was improved with the AUCs of 0.906 and 0.883 in the two cohorts, respectively. This result indicated that both the tumor and peritumoral regions likely reflect tumor genetic information to a certain extent, combining them can enhance the ability to predict the presence of T790M resistance mutation. To better understanding the local variations and heterogeneity within the tumor, we divided the whole ROI into three habitats. The subregion-based models yielded moderate performance, with AUCs ranging from 0.719 to 0.873. Habitat 1 resided in the enhancing tumor core with high cellularity, angiogenesis, and intratumoral heterogeneity, showing the strongest individual predictive ability for T790M mutation. Habitat 3, located at the peritumoral invasive front with subtle microenvironmental changes, presented moderate predictive performance. In comparison, distal peritumoral habitat 2 mainly reflected non-specific edema and gliosis, and thus yielded the lowest predictive value. Notably, the integrated habitat model combining these three distinct subregions outperformed all single region models and achieved the optimal predictive efficacy in this study. It demonstrated that the interaction of different regions of the tumor is also an important factor to predict mutation status. Finally, a habitat radiomics-based nomogram was constructed and performed the best to identify the emergence of T790M mutation after EGFR TKI treatment, outperforming all existing models (Table 3). Subsequently, this study provides an exploratory proteomics explanation for the nomogram. GO enrichment analysis demonstrated preliminary enrichment of pathways involved in RNA processing, RNA splicing, and the spliceosomal complex. These findings showed marked convergence with the spliceosome pathway identified in KEGG analysis, which constitute the most prominent differential module. Collectively, these results offer hypothesis-generating evidence that dysregulated RNA splicing may occur in the high-risk group, which could potentially contribute to T790M resistance mutation through altered splicing isoforms of target genes.

The habitat radiomics-based nomogram may provide supportive evidence to guide diagnostic decisions related to targeted therapy. Firstly, based on conventional imaging, this model can quantitatively assess the risk of T790M resistance mutation, prioritize high-risk patients for liquid biopsy, potentially improve the positive predictive value of detection, and reduce unnecessary testing in low-risk populations. Notably, for patients with negative liquid biopsy but high model-predicted resistance risk, it may alert clinicians to re-evaluate test reliability and reduce diagnostic bias caused by false negatives. Secondly, during routine follow-up imaging, the mode can help capture dynamic changes in resistance‑related imaging features, offer early indications of emerging T790M resistance mutation, and assist clinicians in identifying the optimal timing for repeated biopsy before definite clinical progression, which may help improve biopsy success and positive rates. In addition, this model may serve as a supportive initial screening tool for T790M resistance mutation, aiding clinicians in identifying high-risk populations. It may help optimize detection strategies, reducing diagnostic risks, and enhancing targeted therapy precision, with substantial clinical translational value, thus offering a promising supplementary approach for individualized treatment decisions in NSCLC BM patients.

However, several limitations remain in this study. First, although the model achieved favorable and stable discriminative performance, the mild decrease in PPV from the training cohort to the test cohort suggests mild optimism bias and slight model overfitting, indicating that the predictive performance may have been slightly overestimated during model development. Prospective multicenter validation is warranted to reduce residual optimism bias and improve generalizability. Second, this study was a retrospective study with a relatively small sample size, particularly regarding the tissue samples available for proteomic analysis. Therefore, prospective validation is required to confirm the model’s application value and generalizability, as well as to further elucidate its underlying biological mechanisms. Third, this study only utilized T1-CE images, whereas the combined use of multiple sequences might better reflect the internal information of the tumor. Fourth, only the largest lesion was analyzed in each patient, which may not fully represent inter‑lesion heterogeneity. Future studies should ideally incorporate multiple lesions per patient to better capture intratumoral and inter‑lesion heterogeneity. Finally, the proteomic analysis in this study suggested that RNA splicing mechanisms may be dysregulated in the high-risk group, which may require further confirmation through future analyses, such as spatial transcriptomics.

Conclusions

This study established a habitat model based on tumoral and peritumoral radiomics, which was demonstrated to have predictive value for identifying the presence of T790M resistance mutation in NSCLC BM patients. By combining the habitat model and clinical predictive factors, a habitat radiomics based nomogram was constructed to achieve precision prediction.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The Institutional Review Boards of Beijing Chest Hospital, Capital Medical University and Shandong Cancer Hospital and Institute, Shandong First Medical University approved this retrospective analysis and waived the need for written informed consent.

Data Sharing Statement

Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0244/dss

qims-16-08-627-dss.pdf ^{(111.1KB, pdf)}

DOI: 10.21037/qims-2026-1-0244