Habitat model based on CEUS for noninvasive prediction of EGFR mutation status in peripheral NSCLC

Abstract

Background

To develop and validate a contrast-enhanced ultrasound (CEUS)-based habitat model for noninvasive prediction of epidermal growth factor receptor (EGFR) mutation status in patients with peripheral non-small cell lung cancer (NSCLC).

Methods

This retrospective study included 187 patients with NSCLC confirmed by histopathology from April 2021 to February 2025. All patients underwent CEUS of the lung before biopsy. Patients for whom complete EGFR gene testing results were available were randomly divided into a training set and a test set at a ratio of 8:2. Habitat imaging was used to differentiate the tumor into distinct regions, and then an unsupervised clustering method was used to extract and analyze habitat features to establish Habitat_Model. The Shapley additive explanations (SHAP) method was used to improve the interpretability of the model. In addition, Rad_Model based on radiomics features of tumors was constructed. Finally, a combined model was established by combining the habitat features and clinical–radiological indicators with logistic regression analysis. The predictive performance was evaluated using receiver operating characteristic curve (ROC), calibration, and decision curve analysis (DCA).

Results

The areas under the curve (AUCs) of Habitat_Model and Rad_Model were 0.898 and 0.857 in the training set, and 0.784 and 0.649 in the testing set, respectively. Habitat_Model showed excellent performance. Incorporating clinical–radiological indicators via the combined model slightly improved its performance, resulting in AUC values of 0.904 and 0.812 for the training and testing sets, respectively. The calibration curves and DCA exhibited excellent fit for the combined model, while providing great clinical net benefit.

Conclusions

Our habitat model demonstrates a good capacity for predicting EGFR mutation status in peripheral NSCLC, providing valuable noninvasive reference for clinical pathways on targeted therapy.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12885-026-16416-1.

Background

Lung cancer has become the leading cause of death among all types of cancer worldwide [1]. In China, there are about 1,060,600 new cases of lung cancer each year, along with more than 733,300 deaths from lung cancer. Non-small cell lung cancer (NSCLC) is the most common histological subtype of lung cancer, accounting for more than 85% of cases [2]. The gene most commonly mutated in NSCLC is epidermal growth factor receptor (EGFR), and a recent study showed that EGFR mutation status plays an important role in guiding targeted therapy in NSCLC patients and improving overall survival and quality of life in patients with peripheral NSCLC [3–5]. At present, EGFR mutation status is mainly assessed clinically via pathological tissue biopsy. However, biopsy is invasive, time-consuming, and expensive, and the tumor specimens obtained by biopsy represent only a small part of the tumor tissue. Given that tumors generally exhibit a complex and heterogeneous structure, biopsy cannot provide a comprehensive overview of the tumor’s status [6, 7].

To date, studies on determining EGFR mutation status in NSCLC patients using clinical imaging methods have focused on computed tomography (CT), but CT imaging features are affected by ionizing radiation [8, 9]. Thus, there is an urgent need to establish a more comprehensive and simple diagnostic method to predict EGFR mutation status.

Radiomics can effectively capture quantitative information in tumor imaging data for analysis and reveal internal textural features that cannot be recognized by the naked eye, avoiding reliance on the subjective judgment of doctors [10]. However, conventional radiomics often focuses only on the overall features of lesions, without in-depth analysis of intratumoral heterogeneity. To help to overcome this limitation, the concept of tumor habitats can be used to divide tumors into subregions by identifying grayscale voxels with similar imaging features. For example, Shang et al. [11] extracted features from different habitats to predict the invasiveness of lung adenocarcinoma and developed a habitat model with an area under the curve (AUC) superior to those of other models, indicating the value of habitat radiomics in lung cancer diagnosis and prognosis. In addition, Wang et al. [12] used ultrasound image-based habitat imaging to predict capsule features associated with a high risk of recurrence of parotid pleomorphic adenoma, which was consistently superior to other models in terms of predictive performance. Compared with conventional radiomics, habitat imaging can more objectively and comprehensively mine imaging features characterized by imaging biomarkers in imaging images, and has advantages in visualizing and quantifying intratumoral heterogeneity [13, 14].

Few reports on the radiological features of contrast-enhanced ultrasound (CEUS) habitat subregions to predict EGFR mutation status in NSCLC patients have been published. This study aimed to recognize features of tumor habitat subregions that resemble biological features based on CEUS, and to construct a habitat model for noninvasively predicting EGFR mutation status in NSCLC patients. Such a model should help clinicians make rapid treatment decisions and promote personalized and effective cancer treatment. Finally, we combined clinical–radiological features with CEUS-based habitat features to construct a combined model with enhanced clinical applicability.

Methods

Patients

The institutional review board granted approval for this retrospective investigation and waived the need for informed consent (approval number: 2025-E0709). We retrospectively collected patients with NSCLC confirmed by histopathology from April 2021 to February 2025. A flow chart presenting the process of patient enrollment is shown in Fig. 1.

Fig. 1.

Flow chart of patient enrollment

A total of 406 consecutive patients were analyzed in this study. All patients underwent CEUS of the lung before biopsy and had complete EGFR gene test results. Finally, 187 patients were admitted and arbitrarily divided into a training set and a test set at a ratio of 8:2. The patients were categorized into two groups based on pathological reports: EGFR(+) group (patients with EGFR mutation) and EGFR(−) group (patients without EGFR mutation). The study design and workflow are shown in Fig. 2.

Fig. 2.

(Study design and workflow)

CEUS examination

All patients were examined before biopsy by a senior sonographer(A) with 30 years of experience with CEUS.Prior to the examination, all patients received breathing training and were instructed to maintain breath-holding or shallow, slow breathing during the contrast procedure. The same ultrasound system was used for all patients, namely, LOFIQ E9 (GE Healthcare, Milwaukee, WI, USA), with the following settings: real-time ultrasound contrast, convex array probe model CI-5, probe frequency 3.5–5 MHz, mechanical index 0.08–0.12, and gain 10–15 dB. The ultrasound contrast agent was SonoVue (Bracco Company, Milan, Italy). Routine ultrasound examination was performed first. The largest section of the lesion were selected for imaging mode. The largest section of the lesion were selected for imaging mode. A total of 2.4 mL of contrast medium dissolved in normal saline was rapidly injected through the cubital vein, and then 5 ml of normal saline was used to flush the tube.The patient’s breathing was kept as stable as possible in a fixed position and the sections selected for observation were consistent across all patients. Continuous dynamic observation was performed for 2 min. The data were stored in DICOM format.

Collection of clinical–radiological indicators

The patients’ clinical data were collected through an electronic medical record information system and included the following variables: age, sex, smoking history, pathological type, carcinoembryonic antigen (CEA), cancer antigen 125 (CA125), cytokeratin 19 fragment 21 − 1 (CYFRA21-1), and neuron-specific enolase (NSE).

The features of contrast-enhanced sonography were evaluated by two sonographers (B and C) with more than 8 years of clinical experience, and included the following: maximum diameter, air bronchogram, vascular signs, uniformity of enhancement, degree of enhancement, and perfusion pattern. Quantitative parameter analysis of the contrast-enhanced ultrasound images was performed with the instrument’s Time Intensity Curve (TIC) software. The lung lesion was selected as the region of interest (ROI) to obtain the curve in TIC, after which the following quantitative parameters were extracted: time to peak (TTP), AUC of TIC (Area), and peak intensity (PI). When there was disagreement about the status of these features, discussion was performed until a consensus was reached.

Transpulmonary biopsy and pathology

A fully automatic 18G biopsy gun (ABN-1820; Curaway, Hangzhou, China) was used for targeted biopsy of enhanced areas within lung lesions. Three to four needles were taken. After samples were fixed in 10% formalin, embedded in paraffin, and sectioned, DNA was extracted and sequenced using kits for the direct sequencing of EGFR regions containing common mutations, mainly exon 19 deletion mutation (19del), exon 21 point mutation (21L858R), and exon 20 insertion mutation (exon20). Pathological classification of the samples was performed by experienced pathologists in accordance with the CAP/IASLC/AMP guidelines [15]. EGFR-positive was defined as the presence of exon 18–21 mutation. According to the gene test results, the EGFR-positive group included the following mutation types: 19del, 21L858R, exon20, 19del combined with 21L858R, and 19del combined with exon20.

CEUS acquisition and segmentation

The CEUS video was converted into static images, starting 5 s before the peak phase of lung ultrasound contrast lesions and ending 4 s after it, with acquisition every 1 s. A total of 1870 contrast static images were acquired. A senior sonographer (D) used ITK-SNAP software (version 4.0.1, http://www.itksnap.org). The ROI was delineated along the largest dimension of the lesion.In order to enhance the biological validity, we avoided necrotic areas within the lesion as much as possible. Another senior sonographer (E) also evaluated the ROI, and when there was disagreement in its extent, a third senior sonographer (F) discussed the case with the other two until a consensus was reached. All physicians were blinded to all clinical pathology information.

Spatial tumor habitats from the CEUS

Based on the CEUS images, each lesion was divided into labeled habitat subregions using the K-means clustering algorithm, and Euclidean distances based on voxel and entropy values were used to identify relationships between samples to optimize the clustering process. Average Calinski–Harabasz (CH) scores at different k values were calculated to evaluate the clustering results.

Feature extraction and selection

We extracted habitat features using the PyRadiomics (version 3.0.1) package in Python (version 3.7). The radiological features of each patient were extracted from the ROI of all contrast moment images, including first-order features, shape features, and textural features.The radiomic features were analyzed using the intra-class correlation coefficient (ICC) for two randomly extracted sonographers from 30 cases, and the radiomic features with ICC > 0.75 were retained.

Z-score normalization was performed on all habitat features to avoid overfitting, feature dimensionality reduction was performed using the least absolute shrinkage and selection operator (LASSO) to reduce redundant features, and the penalty parameter (λ) was adjusted according to the nonzero coefficient determined by 10-fold cross-validation and the minimum criterion to determine the habitat features and radiomic features most predictive of EGFR mutation status.

Model performance evaluation

Based on LASSO screening features, Habitat_Model and Rad_Model were constructed by the AdaBoost machine learning algorithm and trained by 10-fold cross-validation to select the optimal parameters to improve their stability.AdaBoost classification algorithm super parameter tuning results are as follows: AdaBoostClassifier (n_estimators = 50, learning_rate = 1.0, random_state = 0).

The clinical–radiological data were analyzed by univariate and multivariate logistic regression, independent correlation factors were screened, and a clinical model was constructed based on independent risk factors using AdaBoost, the same machine learning algorithm as used for the habitat model. Finally, we used a logistic regression (LR) machine learning model to construct a combined model integrating both the habitat model and the clinical model. The predictive performance was compared among Habitat_Model, Clinic_Model, and Combined_Model.

Histologically specific validation cohort analysis

To evaluate whether the original habitat model exploited histological disparities between adenocarcinoma and squamous cell carcinoma as surrogate signals for EGFR mutation prediction, we collected two independent external validation cohorts from the First Affiliated Hospital of Guangxi Medical University (February 2025 – April 2026). All inclusion and exclusion criteria remained identical to those of the original study. Using these data, we constructed two histology-specific validation sets: an adenocarcinoma-only set (AD_Model) and a squamous cell carcinoma-only set (SCC_Model).

Evaluation of model performance

To evaluate the predictive performance of the models, we plotted receiver operating characteristic (ROC) curves and calculated AUC, accuracy, precision, sensitivity, specificity, positive predictive value, negative predictive value, recall, and F1 score. Calibration curves were used to evaluate the goodness of fit of the combined model. When the predictive calibration curve more closely resembled the standard curve, this represented better accuracy of the model. Decision curve analysis (DCA) was also used to quantify net benefit in order to assess the clinical utility of the model at different threshold probabilities. Integrated discrimination improvement (IDI) was used to assess the discriminative power of the model for predictive probability. Here, IDI > 0 was considered to indicate good discriminative power. In addition, the net reclassification index (NRI) was used to quantify the risk stratification of the models’ patient classification. NRI > 0 was considered to indicate improved classification accuracy. Furthermore, CIC was used to assess the impact of the predictive models in actual clinical practice and assess its net clinical benefit. Finally, we used the above five curves to compare and evaluate the models and applied Shapley additive explanations (SHAP) to explore the models’ interpretability.

Statistical analysis

Continuous data are expressed as mean ± standard deviation and were analyzed by t-test or nonparametric rank sum test.Categorical variables are expressed as frequency (%). Count data are expressed as number of cases, analysis of which was performed by chi-square test or Fisher’s exact probability method. The consistency of the results of quantitative parameters of ultrasound evaluation by ICC.The DeLong test was used to analyze the significance of differences in AUC between different models. P < 0.05 was considered statistically significant.Data processing, model construction, statistical analysis, and model evaluation were all carried out using Python (version 3.7), R software (version 4.2.2), and SPSS (version 27.0.1).

Results

Clinical features of the patients

Of the 187 NSCLC patients, 100 were male (53.47%) and 87 were female (46.53%), while their mean age was 61.28 ± 11.09 years (range: 27–93 years). Overall, 118 were included in the EGFR(+) group, while 69 were in the EGFR(−) group. The dataset was randomly divided into a training set and a test set at an 8:2 ratio, with 149 patients in the training group and 38 patients in the test group. Regarding the mutation types, there were 68 patients (36.36%) with EGFR 19del, 40 (21.39%) with 21L858R, 5 (2.67%) with exon20, 2 (1.06%) with 19del with 21L858R, and 3 (1.60%) with 19del with exon20.

There were no significant differences (p > 0.05) in any clinical–radiological parameters of the patients in the training group and the test group within the cohort. Patients’ clinical data and features are presented in Table 1.Two sonographers measured the quantitative parameters of ultrasound contrast, TTP, Area, and PI, which had good stability and reproducibility (ICC = 0.984, 0.828, 0.952, p < 0.001). (Suppl. Table 1) The results of the univariate and multivariate logistic regression analyses in the training set are shown in Table 2. Air bronchogram was the only variable identified by both of these analyses as an independent predictor of EGFR mutation status in NSCLC patients.

Table 1.

Clinical–radiological variables of the patients

Indicator	Total (n = 187)	Training set (n = 149)	Test set (n = 38)	p-value
Age	61.28 ± 11.09	60.62 ± 10.70	63.89 ± 12.32	0.104
Sex				0.765
Male	100 (53.48)	81 (54.36)	19 (50.00)
Female	87 (46.52)	68 (45.64)	19 (50.00)
Smoking history				1.000
No	112 (59.89)	89 (59.73)	23 (60.53)
Yes	75 (40.11)	60 (40.27)	15 (39.47)
Pathological type				1.000
Adenocarcinoma	154 (82.35)	123 (82.55)	31 (81.58)
Squamous cell carcinoma	33 (17.65)	26 (17.45)	7 (18.42)
CEA	254.36 ± 1212.37	290.02 ± 1352.18	114.54 ± 223.87	0.435
CA125	317.02 ± 1067.15	296.27 ± 1067.43	398.39 ± 1076.41	0.076
CYFRA21-1	23.34 ± 57.64	21.91 ± 58.36	28.98 ± 55.12	0.312
NSE	29.31 ± 83.17	21.79 ± 14.28	58.82 ± 181.20	0.494
Maximum meridian	6.40 ± 2.83	6.39 ± 2.81	6.44 ± 2.95	0.976
Enhancement degree				0.567
High	64 (34.22)	49 (32.89)	15 (39.47)
Low	123 (65.78)	100 (67.11)	23 (60.53)
Enhanced mode				0.378
Non-centripetal	75 (40.11)	56 (37.58)	19 (50.00)
Centripetal	53 (28.34)	44 (29.53)	9 (23.68)
Mixed	59 (31.55)	49 (32.89)	10 (26.32)
Uniformity				1.000
No	164 (87.70)	131 (87.92)	33 (86.84)
Yes	23 (12.30)	18 (12.08)	5 (13.16)
Air bronchogram				0.296
No	44 (23.53)	38 (25.50)	6 (15.79)
Yes	143 (76.47)	111 (74.50)	32 (84.21)
Vascular signs				0.630
Branch sign	55 (29.41)	44 (29.53)	11 (28.95)
Curly hair sign	39 (20.86)	29 (19.46)	10 (26.32)
None	93 (49.73)	76 (51.01)	17 (44.74)
TTP	34.51 ± 27.23	35.09 ± 28.45	32.24 ± 21.94	0.721
Area	1450.67 ± 1388.38	1445.43 ± 1513.64	1471.23 ± 724.63	0.328
PI	20.41 ± 7.28	20.55 ± 5.99	19.85 ± 11.05	0.838

Table 2.

Univariable and multivariable analyses of clinical features

Indicator	Univariate logistic regression			Multivariate logistic regression
	OR (95% CI)		p value	OR (95% CI)	p value
Age		1.008 (1.003–1.012)	0.006	0.993 (0.965–1.021)	0.668
Sex	1.269 (1.070–1.505)		0.021	1.022 (0.438–2.385)	0.966
Smoking history	1.256 (1.044–1.511)		0.043	0.419 (0.178–0.982)	0.093
Pathological type	1.319 (1.055–1.650)		0.041	0.41 (0.177–0.946)	0.08
CEA	1.001 (1.000–1.001)		0.163
CA125	1.000 (1.000–1.000)		0.840
CYFRA21-1	1.005 (0.999–1.011)		0.177
NSE	1.021 (1.009–1.033)		0.003	1.013 (0.989–1.037)	0.386
Maximum meridian	1.048 (1.007–1.09)		0.053
Enhancement degree	1.284 (1.094–1.505)		0.010	0.885 (0.438–1.790)	0.776
Enhanced mode	1.309 (1.143–1.498)		0.001	1.532 (1.061–2.214)	0.056
Uniformity	1.292 (1.116–1.495)		0.004	0.858 (0.352–2.088)	0.777
Air bronchogram	1.424 (1.215–1.669)		0.000	4.648 (2.151–10.044)	0.001
Vascular signs	1.358 (1.163–1.586)		0.001	1.169 (0.772–1.770)	0.537
Ttp	1.011 (1.004–1.017)		0.009	1.000 (0.989–1.012)	0.999
Area	1.000 (1.000–1.000)		0.245
PI	1.025 (1.012–1.039)		0.002	0.986 (0.931–1.044)	0.675

Based on the CH score (Suppl. Figure. 1), the optimal number of clusters for the K-means algorithm was determined to be 3.This means that three habitat subregions were generated, indicating tighter clustering, greater differences between clusters, and improved performance. Therefore, the ROI of tumor was separated into three habitat subregions for subsequent feature extraction and model construction.

Conventional radiomics analysis generated 105 features. In addition, we extracted features from three different subregions to analyze habitat. We extracted 107 habitat features from each habitat subregion of each CEUS image and averaged them individually across all habitat subregions of the same lesion.After averaging, a total of 321 features were extracted from each habitat subregion of each lesion. After feature selection, 13 features were chosen for subsequent model construction of the habitat region.

Model performance

The AUCs of Habitat_Model and Rad_Model were 0.898 and 0.857 in the training set, and 0.784 and 0.649 in the testing set, respectively. While the difference in AUC between the Habitat_Model and Rad_Model was not statistically significant (p > 0.05), DCA (Suppl. Figure. 2) revealed that the Habitat_Model offered a greater net clinical benefit compared to the Rad_Model within the clinically relevant threshold probability range of 0.1–0.8. Table 3 summarizes the predictive performance of all models. Combined_Model consistently demonstrated excellent predictive performance across all datasets.

Table 3.

Comparison of the performance of the different models

Model	Set	AUC (95% CI)	ACC	SEN		SPE	NPV	PPV	PRE	Recall	F1 score
Habitat_Model	Train	0.898 (0.850–0.946)	0.812		0.766	0.891	0.690	0.923	0.923	0.766	0.837
Habitat_Model	Test	0.784 (0.629–0.939)	0.763		0.708	0.857	0.632	0.895	0.895	0.708	0.791
Clinic_Model	Train	0.600 (0.525–0.676)	0.658		0.819	0.382	0.553	0.694	0.694	0.819	0.751
Clinic_Model	Test	0.601 (0.466–0.736)	0.684		0.917	0.286	0.667	0.687	0.687	0.917	0.786
Combined_Model	Train	0.904 (0.858–0.951)	0.805		0.734	0.927	0.671	0.945	0.945	0.734	0.826
Combined_Model	Test	0.812 (0.667–0.958)	0.789		0.708	0.929	0.65	0.944	0.944	0.708	0.810

Abbreviations: AUC Area under the curve, 95% CI 95% confidence interval, ACC Accuracy, SEN Sensitivity, SPE Specificity, NPV Negative predictive value, PPV Positive predictive value, PRE precision

We compared the AUC values for all models based on the above results to make the performance comparisons more intuitive (Suppl. Figure. 3). The calibration curve (Suppl. Figure. 4) showed that the bias curves for the training and test sets of the combined model were close to ideal, suggesting that the combined model had good predictive accuracy and power. Moreover, DCA (Suppl. Figure. 5) indicated that the use of the combined model to predict EGFR mutation status in NSCLC would be better than using full treatment (i.e., assuming that all lesions are positive for EGFR mutation) or no treatment (i.e., assuming that all lesions are negative for EGFR mutation). CIC (Suppl. Figure. 6) showed that the population predicted to have EGFR mutations by the combined model was likely to strongly overlap with the actual EGFR-mutation-positive population above a risk threshold of 0.4.

IDI (Suppl. Figure. 7) suggested that the combined model improved the predictive discrimination of EGFR mutations. Moreover, NRI (Suppl. Figure. 8) suggested that the combined model improved the classification accuracy by correctly reclassifying patients into the correct risk stratum.

Figure 3 demonstrates the practical application of our habitat model in the clinic, which has significant value in pre-biopsy screening or guiding biopsy targeting.

Fig. 3.

Comparison of traditional approach based on clinical guidelines and prediction process with Habitat_Model. a Traditional approach based on clinical guidelines for selecting lung cancer treatment. This process includes the following main steps: assessment of the malignancy of lung tumor through ultrasound and clinical information; determination of whether to perform ultrasound-guided biopsy of tumor; performance of pathological examination on successfully biopsied tissues; and formulation of treatment plan by determining the status of EGFR gene mutation based on pathological results. b Clinical guidelines for treatment planning regarding status of EGFR gene mutation in lung cancer combined with Habitat_Model. Habitat_Model assists in the direct noninvasive prediction of EGFR gene mutation status and treatment planning

Histologically specific validation cohort

The adenocarcinoma validation set included 33 patients with lung adenocarcinoma, comprising 18 EGFR non-mutant and 15 EGFR mutant cases. The squamous cell carcinoma validation set included 15 patients with lung squamous cell carcinoma, comprising 10 EGFR non-mutant and 5 EGFR mutant cases. As shown in Suppl. Table 2, the Habitat_Model achieved an AUC of 0.755 (95% CI: 0.592–0.918) in AD_Model and 0.720 (95% CI: 0.345–1.000) in SCC_Model.

Graphical interpretation of the habitat model

To explain the influence and relative contributions of habitat features to the prediction of EGFR mutation status, we used SHAP to visualize the prediction process of Combined_Model. As expected, habitat features were identified as important contributors to such predictions, as shown in a SHAP summary bar plot (Figure. 4a). A SHAP beeswarm plot (Figure. 4b) showing the positive or negative influence of each feature on the probability of positivity for EGFR mutation is also presented. Finally, SHAP heat maps are provided in Figure. 5, showing two clinical cases that were correctly predicted as EGFR(−) and EGFR(+).

Fig. 4.

Importance of SHAP analysis variables for each index of the habitat model. a The SHAP summary bar plot shows the order of importance of the habitat features, with the most important features appearing at the top. b The SHAP beeswarm plot shows each feature’s positive or negative effects on the probability of positivity for EGFR mutation in different colors (red corresponds to high feature values and blue corresponds to low ones)

Fig. 5.

Typical case display. Panels a–f show an EGFR(−) patient (male, 73 years old), while d–f show an EGFR(+) patient (female, 66 years old). a and d Contrast ultrasound images. b and e Pathological indication of lung adenocarcinoma (HE ×400). c. The prediction basis E[f(x)] of this case is larger than the output f(x) of this individual. The “original_glszm_HighGrayLevelZoneEmphasis_h2” feature has a negative influence on prediction of 3.268, and the color region width of this feature is the largest. This indicates that this feature has the greatest influence on the probability of positivity for EGFR mutation in this patient. This case is correctly predicted as negative for EGFR mutation. f The prediction base value of this case is less than the output value f(x) of this individual, and the “original_glszm_SmallAreaHighGrayLevelEmphasis_h3” feature has a positive influence on the prediction of 1.335. The color region width of this feature is the largest. This indicates that this feature has the greatest influence on the probability of positivity for EGFR mutation in this patient. This case is correctly predicted as positive for EGFR mutation.Red represents a positive contribution and blue a negative one. Base value represents the basic prediction probability of the model, while f(x) represents the final prediction probability

Discussion

Our study involved the construction of a habitat feature model based on CEUS intratumoral habitat subregion clustering to predict EGFR mutation status, with AUCs (95% CI) of 0.898 (0.850–0.946) and 0.784 (0.629–0.939) in the training and test sets, respectively. Thus, our model showed better predictive performance than conventional radiomics models. Conventional radiomics analysis involves the extraction of quantitative features of the whole tumor from CEUS images, which may overlook intratumoral heterogeneity.

To overcome this limitation, habitat analysis has been developed as a way of capturing, Habitat analysis is a new method that can capture subtle differences within tumors [16]. To characterize intratumoral heterogeneity more effectively, habitat imageomics was introduced into model construction in this study. Specifically, each tumor was divided into multiple subregions using voxel clustering, which enabled different cell subsets to be identified. In our study, three different habitat subregions were identified by optimizing the number of clusters via K-means clustering. At the microscopic level, these subregions may reflect the spatial distribution of clusters of cells with similar biotypes.By further exploring the distribution of tumor cells and their molecular properties, intra-tumor heterogeneity may be better quantified [17]. To enhance the model’s reliability in handling unseen data and address potential overfitting, we implemented a rigorous feature selection and dimensionality reduction process. This involved initial screening based on statistical tests to retain features significantly associated with EGFR mutation status, followed by Spearman correlation analysis to remove highly collinear features. Finally, LASSO regression was applied to further compress the feature space and retain only the most predictive subset. These steps aimed to reduce overfitting risk, improve generalization, and mitigate the performance gap between training and testing sets.As expected, the subregion-based habitat model provides valuable information for predicting EGFR mutation status in NSCLC. Sujit et al. [18] used habitat imaging based on CT and FDG-PET scan images to effectively stratify NSCLC patients according to the level of risk of recurrence after surgery, highlighting the potential of habitat analysis-based approaches. Meanwhile, the habitat features extracted by Wu et al. [19] were shown to be promising in distinguishing EGFR mutants and wild type. Habitat imaging has also been confirmed to deepen our understanding of tumor heterogeneity [20], while providing significant advantages in the use of key molecular markers of tumors for predictive and diagnostic purposes.

To enable a more intuitive and interpretive visualization of the habitat model, SHAP values [21] were also used to calculate the contribution of each feature in the model to the prediction of EGFR mutation status. In the ranking list of importance of habitat features (Figure 4), we observe that“original_ngtdm_Strength_h2”had the highest ranking and was scattered. This feature affects the model output to some extent, and the color displayed shows that higher values for this feature correlate with increased model output.This indicates that this feature not only makes a higher overall contribution but is also the most important of all features.Different habitats may correspond to different subregions with distinct biological behaviors.This habitat feature, derived from the Neighborhood Gray Tone Difference Matrix (NGTDM), quantifies the intensity of local gray-level variation within the image [22]. Higher values of this feature correlate with an increased probability of predicting EGFR mutation by the model. From a biological perspective, EGFR-mutant tumors often exhibit heightened heterogeneity in cellular proliferation. Variations in cell density and proliferative activity across different tumor regions manifest as localized fluctuations in gray-level values on CEUS images. The NGTDM texture feature effectively captures this intensity of gray-level variation, thereby serving as an indirect reflection of the underlying molecular biological characteristics within the tumor. Xu et al. [23]predicted EGFR expression status in breast cancer tumors based on ultrasound imaging. As one of the top predictors selected for in the radiomics model, “original_ngtdm_Strength_h2”makes certain contributions in helping to predict EGFR mutations. An interpretable analysis key principal component features indicates that certain high-weight habitat features could potentially be closely associated with molecular or pathological characteristics of tumor infiltration and mutation, thereby offering insights into their biological mechanisms for further investigation.SHAP Force Plot was used in this study to assess the EGFR mutation status of individual patients, and the habitat model can determine their risk level. The SHAP-based interpretation of individual sample predictions also clearly indicates the specific contribution of each habitat feature to individual sample predictions, which improves the transparency of habitat models and their value for application in clinical practice.

Another innovation of this study is the extraction of habitat features from CEUS images. CEUS examination has become the tool of choice for assessing peripheral lung tumors because of its noninvasiveness and because it is widely available. However, ultrasound can only clearly visualize lesions close to the pleura, and the features reflective of EGFR mutation status are not always clear [24, 25]. While CT holds an irreplaceable role in lung cancer diagnosis, with its excellent spatial resolution and widely established global standardization well-validated.It has been shown to be valuable in predicting EGFR mutation status [26]. However, the core advantage of CEUS lies in its ability to provide functional information on tumor microvascular perfusion without radiation, in real-time, and dynamically. It can be safely used in patients with renal insufficiency and intuitively visualizes the heterogeneous spatial patterns of intratumoral blood flow perfusion [27]. Thereby, it adds a unique functional dimension to the existing diagnostic pathway, particularly suitable for clinical scenarios where contrast-enhanced CT is contraindicated or immediate hemodynamic assessment is required. In this study, we used CEUS combined with habitat imaging to predict EGFR mutation status. The habitat model obtained satisfactory results, highlighting its potential as an effective tool for predicting EGFR mutation status. Therefore, it has broad prospects for application in the clinical evaluation of NSCLC patients.

Given the predominance of adenocarcinoma in the original cohort and the higher prevalence of EGFR mutations in this histologic subtype, the habitat model may have partially leveraged imaging differences between adenocarcinoma and squamous cell carcinoma as surrogate features for EGFR mutation prediction, rather than capturing true mutation-specific imaging phenotypes. To test this hypothesis, we constructed two histologically homogeneous external validation sets consisting of pure adenocarcinoma and pure squamous cell carcinoma, respectively. The results confirmed that histologic confounding indeed influenced model performance; however, the habitat model retained a certain degree of discriminative ability for EGFR mutations within a single histologic context.Furthermore, the habitat model maintained independent predictive value specifically in lung adenocarcinoma, the most clinically relevant population.

Our study shows that EGFR mutations are independently associated with air bronchogram. Consistent with the findings of most studies on the association of clinical and imaging features with EGFR mutations in lung cancer [28, 29], positivity for EGFR mutation in this study was more common in patients with air bronchogram in the lesions. This may be explained by the spread of the air cavity to the periphery in the presence of the gene mutation, which would occupy the adjacent bronchioles and alveolar space. This would in turn obstruct lung parenchyma around the tumor or obstruct terminal bronchioles, and then cause lung consolidation.Although the use of preoperative clinical features has progressed, these indicators need to be manually characterized or measured. Accurate judgment of air bronchial signs requires high physician level and is prone to errors. Its predictive power is still limited

Finally, we integrated the clinical features with CEUS-based habitat features to construct a combined model with enhanced clinical applicability. This combined model not only showed excellent performance in the training set (AUC: 0.904) but also maintained good performance in the test set (AUC: 0.812). The AUC of the combined model was better than those of the habitat model and clinical model in both training and test sets. The SHAP swarm plot of the combined model( Suppl. Figure. 9 ) visually demonstrates the extent and direction of habitat and clinical features contributing to predicting EGFR mutation status.The combined model thus effectively solves the limitation of insufficient sensitivity of the clinical model by combining habitat with clinical index. We further evaluated the consistency of the prediction probability and observation probability of the combined model by using a calibration curve. The results showed that the calibration curve fitted the standard curve well, indicating that the EGFR mutation status predicted by the combined model is consistent with the actual mutation status and that the model has good consistency and predictive accuracy. Notably, the combined model also showed high clinical net benefit, suggesting that its superior performance and highly consistent predictions provide doctors with a practical tool for noninvasively assessing EGFR mutation status in peripheral NSCLC, which can facilitate clinical treatment decisions.

Despite the strengths of this work, there are some limitations that should be mentioned. First, this study focused on differentiation of EGFR mutation and wild-type status in all NSCLC patients, without assessing subtype specificity and subgroup analysis of adenocarcinoma populations. Future studies need to analyze the expression of EGFR mutation subtypes in different populations to improve clinical relevance.Second, EGFR has spatial heterogeneity, and the biopsy interface and ultrasound contrast interface cannot be completely matched. Third, our work involved a single-center, small-sample retrospective study. It would potentially overfit or reduce power for subgroup analyses.Therefore, future multicenter studies should employ ultrasound devices from diverse brands and models for data acquisition, and explore techniques such as federated learning to integrate these data, thereby further enhancing the performance and generalizability of the predictive model.Finally, the “black box” effects inherent in habitat models leave the biological significance behind the signature unclear. we plan to use spatial transcriptome technology for further prospective studies.

Conclusions

In conclusion, the habitat radiomics analysis strategy that we developed in this study based on CEUS demonstrated excellent performance in predicting the EGFR mutation status of peripheral NSCLC patients.

Abbreviations

ACC

Accuracy

Area

AUC of TIC

AUC

Areas under the curve

CEUS

Contrast-enhanced ultrasound

CIC

Clinical impact curve

DCA

Decision curve analysis

EGFR

Epidermal growth factor receptor

ICC

Intra-class correlation coefhcient

IDI

Integrated discrimination improvement

LASSO

Least absolute shrinkage and selection operator

LR

Logistic regression

NPV

Negative predictive value

NRI

Net reclassification index

NSCLC

Non-small cell lung cancer

NGTDM

Neighborhood Gray Tone Difference Matrix

PI

Peak intensity

PPV

Positive predictive value

PRE

Precision

ROI

Region of interest

SEN

Sensitivity

SHAP

Shapley additive explanations

SPE

Specificity

TIC

Time intensity curve

TTP

Time to peak

19del

Exon 19 deletion mutation

exon20

Exon 20 insertion mutation

21L858R

Exon 21 point mutation

Authors' contributions

J.Z. conceived and designed the study, developed the methodology, performed validation, conducted the investigation, curated the data, and wrote the original draft. J.Z. also participated in reviewing and editing the manuscript. Y.G. and X.L. contributed to the conceptualization of the study, provided resources, and participated in reviewing and editing the manuscript. L.W. contributed to the development of the methodology and performed the formal analysis. H.C. developed the software and participated in the validation process. Y.L. contributed to the investigation and created the visualizations. F.H. curated the data and supervised aspects of the research. T.Q. acquired the funding and administered the project. All authors have read and agreed to the published version of the manuscript.All authors read and approved the final manuscript.

Funding

This research received no external funding.

Data availability

Due to ethical restrictions, the raw data cannot be made publicly available. However, all data may be obtained from the corresponding author upon reasonable request.

Declarations

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. We all agree to submit it for publication and are aware of and accept the journal's publication policies.

Ethics approval and consent to participate

The study was conducted in accordance with the Declaration of Helsinki, and approved by the Ethics Review Committee of First Affiliated Hospital of Guangxi Medical University (protocol code 2025-E0709 and date of approval: August 27, 2025).Informed consent has been obtained from all participants in this study. All the listed authors have carefully reviewed and approved this manuscript. We all agree to submit it for publication and are aware of and accept the journal’s publication policies.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.