Differentiating breast cancer lung metastasis from primary lung cancer using clinical and computed tomography radiomics features

Introduction

Breast cancer is one of the most prevalent malignancies among women worldwide and remains a leading cause of cancer-related morbidity and mortality, largely due to metastatic disease (1). Among breast cancer survivors, breast cancer lung metastasis (BCLM) and primary lung cancer (PLC) represent distinct disease entities with substantially different prognoses and therapeutic approaches, creating significant diagnostic challenges in clinical practice (2,3).

Pulmonary nodules detected by computed tomography (CT) in patients with breast cancer are frequently presumed to represent metastatic disease at initial assessment. However, pathological confirmation has demonstrated that approximately 48% to 67% of these nodules are PLC, whereas only 23% to 43% are BCLM (4). This diagnostic discrepancy is partly attributable to the routine use of follow-up CT imaging and the overlap in radiological features between PLC and BCLM, including irregular margins and central defects (5). Diagnostic uncertainty is further increased in nodules smaller than 1 cm, as subtle imaging characteristics limit reliable differentiation, highlighting the need for a non-invasive and accurate diagnostic approach.

Prior studies have reported that patients with PLC are generally older, present with larger pulmonary nodules, and exhibit higher estrogen receptor (ER) positivity compared with patients with BCLM, who more frequently demonstrate lymph node positivity and a higher prevalence of triple-negative breast cancer (6). Zhao et al. reported longer disease-free intervals among patients with BCLM, indicating more advanced underlying disease stages compared to those with PLC (7). Also, He et al. identified age, tumor size, and prior radiotherapy as independent predictors of secondary PLC in an analysis of 55 patients with PLC and 205 patients with BCLM (8). Although these findings provide valuable clinical insights, most prior studies were limited in terms of small sample size and did not establish a robust predictive model for accurate differentiation between PLC and BCLM.

Radiomics refers to the extraction of high-dimensional quantitative features from medical images or predefined regions of interest (ROIs). These features characterize tissue heterogeneity, intensity distribution, and morphological properties and can be integrated with clinical, pathological, genomic, or proteomic data to address complex diagnostic and prognostic questions (9). In recent years, radiomics has been increasingly applied to lung tumors for diagnosis, molecular subtype classification, prediction of treatment response, and prognosis assessment (10-12).

In this retrospective analysis, clinical characteristics of patients with BCLM and PLC were compared, and radiomics features were extracted from CT-defined ROIs in lung nodules. The objective was to develop a highly accurate predictive model to improve diagnostic discrimination of pulmonary nodules in patients with breast cancer, thereby supporting more precise clinical decision making and individualized treatment strategies. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0600/rc).

Methods

Data source and study population

Patients diagnosed with breast cancer and concomitant PLC or BCLM at the Affiliated Union Hospital of Fujian Medical University between 2013 and 2023 were retrospectively included. Clinical, pathological, and CT imaging data were collected. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of the Affiliated Union Hospital of Fujian Medical University (No. 2025KY029). Informed consent was taken from all the patients (or a statement that it was not required and why).

The inclusion and exclusion criteria were defined as follows.

Inclusion criteria

(I) Patients who were treated at the Affiliated Union Hospital of Fujian Medical University between January 2013 and October 2023. (II) Pathological confirmation of breast cancer. (III) Pathological confirmation of PLC or BCLM, with lung lesion diagnosis occurring within 1 year of breast cancer diagnosis. (IV) Availability of complete clinical, imaging, and pathological data obtained before and during treatment.

Exclusion criteria

(I) Lung nodules diagnosed more than one year after the initial diagnosis of breast cancer. (II) Incomplete key clinical data, including age, menopausal status, laterality of breast cancer, tumor location, tumor (T) stage, node (N) stage, ER status, progesterone receptor (PR) status, human epidermal growth factor receptor 2 (HER2) status, Ki-67 index, luminal subtype, and serum levels of carcinoembryonic antigen (CEA), carbohydrate antigen 125 (CA125), carbohydrate antigen 153 (CA153), or other essential clinicopathological parameters. (III) Poor-quality or incomplete CT imaging data. (IV) Male patients with breast cancer. (V) Patients presenting with distant metastasis at initial diagnosis or with a history of malignancies other than breast cancer or lung cancer. (VI) History of chest radiotherapy prior to CT image acquisition.

Clinical model construction

Clinical data collection

Clinical data were collected, including age at diagnosis, laterality of breast cancer (left, right, or bilateral), tumor location (upper inner quadrant, lower inner quadrant, upper outer quadrant, lower outer quadrant, or central region), and tumor size and extent as defined by T stage. T stage was classified from T1 to T4 according to tumor size and local invasion. Lymph node involvement was classified as N0 for no involvement and N1 to N3 for increasing extent of nodal involvement. Serum tumor markers, including CEA, CA125, and CA153, were measured at the time of breast cancer diagnosis. Elevated levels were defined as CEA >5 ng/L, CA125 >35 U/mL, and CA153 >25 U/mL and were categorized as high or low based on established clinical thresholds (13).

Immunohistochemistry (IHC) results for molecular markers ER, PR, HER2, and Ki-67 were reported by the pathology department of the hospital. ER and PR were considered positive if nuclear staining exceeded 1%, and HER2 was considered positive if more than 10% of tumor cells demonstrated strong, uniform membrane staining. For Ki-67, high expression was defined as ≥20%, and low expression as <20%. From the IHC results, breast cancer luminal subtypes were classified as hormone receptor (HR)+/HER2−, HER2+, and HR−/HER2−.

Clinical data analysis and prediction model construction

Patients were randomly assigned to training and testing sets in a 70% to 30% ratio using R software. Univariate logistic regression analysis was conducted to screen clinicopathological variables, and factors with P<0.1 were entered into multivariate logistic regression analysis. Variables with P<0.05 were considered independent predictors and were incorporated into the clinical prediction model. Model discrimination was evaluated using the concordance index (14). Calibration curves were generated using bootstrap resampling to assess agreement between predicted probabilities and observed outcomes. Receiver operating characteristic (ROC) curves were constructed to assess model performance.

Imaging model construction

CT imaging acquisition

All patients underwent chest CT following breast cancer diagnosis using a GE Revolution 256 slice CT scanner. Imaging parameters included a tube voltage of 120 kVp, automatic tube current modulation, a field of view of 320 mm × 320 mm, and slice thickness ranging from 1.25 to 5.0 mm. Lung and mediastinal window images were retrieved in Digital Imaging and Communications in Medicine format from the Picture Archiving and Communication System and converted to Neuroimaging Informatics Technology Initiative format using SimpleITK (Python version 3.6.13). To reduce variability related to slice thickness, images were resampled to an isotropic voxel size of 1 mm × 1 mm × 1 mm. The first scan demonstrating lung nodules was selected for analysis when multiple CT scans were available.

CT image segmentation and feature extraction

CT images were independently segmented by two experienced radiologists using three-dimensional (3D) Slicer version 5.4.0 (https://www.slicer.org). Tumors were delineated on lung window images to define ROIs. The tumor margin was expanded by 5 mm to generate the peritumoral region (PTR). All segmentations were subsequently reviewed and confirmed by senior radiologists to ensure accuracy and consistency. Radiomics features from intratumoral and PTRs were extracted using the Radiomics plugin in 3D Slicer following wavelet preprocessing. Reproducibility was evaluated by repeat segmentation of 35 randomly selected cases after a 30-day interval. Features with an intraclass correlation coefficient (ICC) >0.75 were considered reliable and were retained for further analysis. The extracted radiomics features included:

• Morphological features: representing two dimensional and 3D tumor characteristics, including volume, surface area, density, and sphericity.
• First-order statistics: these features described the distribution of voxel intensity values within the ROIs, including mean, standard deviation, skewness, kurtosis, and related metrics.
• Texture features: these features quantified tumor heterogeneity and inter-voxel spatial relationships, including gray-level dependence matrix (GLDM), gray-level co-occurrence matrix (GLCM), gray-level size zone matrix (GLSZM), gray-level run length matrix (GLRLM), and neighborhood gray-tone difference matrix (NGTDM).
• Wavelet features: these features were derived from wavelet decomposition of the ROIs along the x, y, and z directions into low-frequency (L) and high-frequency (H) components, resulting in eight wavelet-transformed feature sets.

Feature selection and model construction

Radiomics features were standardized using Z score normalization. Independent t-tests were conducted to identify features with statistically significant differences between groups, with P<0.05 considered significant. Least absolute shrinkage and selection operator (LASSO) regression analysis was used to reduce dimensionality and select the most informative features. Tenfold cross validation was used to minimize model error and determine the optimal regularization parameter λ. The selected features were weighted according to their corresponding regression coefficients to calculate the radiomics score (Rad score). Logistic regression analysis was subsequently used to construct radiomics prediction models based on intratumoral region (ITR), PTR, and multi region (MR) features.

Combined model construction and validation

Independent clinical predictors identified through multivariate logistic regression analysis and radiomics scores derived from the radiomics models were integrated to develop a combined prediction model. Model performance was assessed by calculating sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and area under the curve (AUC) with 95% confidence intervals (CIs) for the clinical, radiomics, and combined models in both the training and testing sets. Calibration curves and concordance index values were used to assess agreement between predicted and observed outcomes. Decision curve analysis (DCA) was conducted to assess the potential clinical use of the models.

Statistical analysis

Statistical analyses were conducted using Python (version 3.11.3) and R (version 4.3.2). The SciPy package was used for feature independence testing, Levene test for assessment of variance homogeneity, and Welch t-test when variance assumptions were not met. Z score normalization was performed using the scikit learn package. R packages including caret, glmnet, pROC, rms, and rmda were used for random data partitioning, LASSO regression analysis, ROC curve analysis, calibration curve construction, and DCA. All statistical tests were two sided, and P<0.05 was considered statistically significant. 95% CIs were reported where appropriate.

Results

Clinical characteristics of patients

A total of 158 patients with breast cancer were included according to the predefined inclusion and exclusion criteria (Figure 1). The cohort was divided into a PLC group (n=100) and a BCLM group (n=58). Clinical characteristics of the two groups are summarized in Table 1. When compared to the PLC group, the BCLM group demonstrated a significantly higher T stage (P=0.004), with a greater proportion of patients classified as T3–T4 (20.7% vs. 7.0%) and a lower proportion classified as T1 (19.0% vs. 40.0%). N stage was also higher in the BCLM group (P<0.001), with increased proportions of patients at N1 (51.7% vs. 19.0%) and N2–N3 (6.9% vs. 1.0%). Also, the BCLM group demonstrated higher HER2 positivity (44.8% vs. 16.0%, P<0.001) and a higher rate of Ki-67 overexpression (82.8% vs. 60.0%, P=0.005). Regarding molecular subtypes, patients in the BCLM group had a higher proportion of the HER2+ subtype (24.1% vs. 10.0%), whereas patients in the PLC group had a higher proportion of the HR+/HER2− subtype (75.0% vs. 58.6%) (P=0.041). Elevated serum CA153 levels were more frequently observed in the BCLM group than in the PLC group (36.2% vs. 20.0%, P=0.04). No statistically significant differences were found between the two groups with respect to age (P=0.07), menopausal status (P=0.92), laterality of breast cancer (P=0.84), tumor location (P=0.74), ER status (P=0.06), PR status (P=0.32), CEA levels (P=0.36), or CA125 levels (P>0.99).

Figure 1 Flow chart of case selection (A) and image segmentation of radiomics (B). BCLM, breast cancer lung metastasis; CT, computed tomography; ITR, intratumoral region; PLC, primary lung cancer; PTR, peritumoral region.

The 158 patients were randomly allocated to training and testing sets in a 7:3 ratio. The training set comprised of 111 patients, including 69 patients with PLC (62.2%) and 42 patients with BCLM (37.8%). The testing set included 47 patients, of whom 31 had PLC (66.0%) and 16 had BCLM (34.0%). No significant difference in lung nodule pathology was observed between the training and testing sets (P=0.78). Comparisons of clinical characteristics between the training and testing sets are presented in Table 2. No statistically significant differences were identified for age at diagnosis, menopausal status, laterality of breast cancer, tumor location, T stage, N stage, ER status, PR status, HER2 status, Ki-67 expression, luminal subtype, or serum levels of CEA, CA125, and CA153 (all P>0.05).

Clinical prediction model construction

Univariate and multivariate logistic regression analyses were conducted in the training set (n=111) to identify independent factors for differentiating between PLC and BCLM. In univariate logistic regression analysis, variables demonstrating potential associations (P<0.1) included T stage, N stage, ER status, HER2 status, Ki-67 expression, luminal subtype, and serum CA153. Since luminal subtype incorporates both HER2 status and Ki-67 expression, the latter two variables were excluded from the multivariate analysis to avoid multicollinearity. Multivariate logistic regression analysis identified N stage and luminal subtype as independent predictors for the clinical prediction model used to distinguish between PLC and BCLM (Table 3, P<0.05). The clinical model was validated in the testing set, yielding a concordance index of 0.760 (95% CI: 0.719–0.801). Model discrimination was further assessed using ROC analysis. In the training set, the AUC was 0.704 (95% CI: 0.606–0.802), with a sensitivity of 64.3% and an accuracy of 56.3%. In the testing set, the AUC was 0.760 (95% CI: 0.615–0.905), with a sensitivity of 75.0% and an accuracy of 74.5%. Calibration curves for both the training and testing sets demonstrated slopes close to 45°, indicating good agreement between predicted probabilities and observed outcomes. ROC and calibration curves for the clinical prediction model are presented in Figure 2.

Figure 2 ROC analysis and calibration plot for clinical model. The ROC curve in the training group (A) and testing group (B), and the calibration plot in the training group (C) and testing group (D). AUC, area under the curve; ROC, receiver operating characteristic.

Radiomics feature model construction

Radiomics features were extracted from ITR and PTR images of all 158 patients using 3D Slicer software. A total of 851 features were obtained from each region, resulting in 1,702 radiomics features per patient. These included 14 morphological features, 18 first-order histogram features, 24 GLCM features, 14 GLDM features, 16 GLRLM features, 16 GLSZM features, and 5 neighboring gray tone difference matrix features. Wavelet transformation was applied to each feature, yielding the complete set of 1,702 radiomics features per case.

Feature reproducibility was evaluated using the pingouin package in Python to assess interobserver and intraobserver consistency, and features with an ICC >0.75 were retained. This process resulted in 1035 stable radiomics features. Subsequently, two-sample t-tests were performed on features from the training and testing sets, and 404 features were selected, including 251 ITR features and 153 PTR features.

Multiple region radiomic (MR_RAD) model

In the training set (n=111), the 404 selected radiomics features underwent dimensionality reduction using LASSO regression with 10-fold cross validation. The regularization parameter λ was adjusted to reduce the mean squared error, with lower values indicating improved model fit. The optimal λ was selected using the 1 standard error criterion, resulting in 10 nonzero features for construction of the radiomics model (Figure 3A,3B). Figure 3A depicts the dimensionality reduction path, with the vertical line indicating the λ value selected using the 1 standard error criterion. Figure 3B presents the coefficient convergence plot, in which the left and right vertical lines correspond to the minimum criterion and the 1 standard error criterion, respectively.

Figure 3 Radiomics feature selection using the LASSO regression in MR_RAD model (A,B), ITR_RAD model (C,D), and PTR_RAD model (E,F). ITR_RAD, intratumoral radiomic; LASSO, least absolute shrinkage and selection operator; MR_RAD, multiple region radiomic; PTR_RAD, peritumoral radiomic.

The MR_RAD model is comprised of three peritumoral features and seven intratumoral features, including one morphological feature, one first-order histogram feature, and eight wavelet-transformed features. The final radiomics score was calculated by weighting each selected feature according to its corresponding regression analysis coefficient, using the following formula: Rad-score = −0.623723689 + 0.153460962 × original.shapeSphericity (ITR) + 0.420923323 × original.firstorderMedian (ITR) − 0.038003416 × wavelet.LHL_firstorderMean (ITR) − 0.171571108 × wavelet.HLL_firstorderMean (ITR) + 0.367039628 × wavelet.HLL_glcm_MCC (ITR) + 0.302213172 × wavelet.HLL_ngtdm_Strength (ITR) + 0.150149958 × wavelet.LLL_glcm_DifferenceVariance (ITR) − 0.058492547 × wavelet.LHL_glcm_Imc1 (PTR) + 0.018413060 × wavelet.LHL_gldm_SmallDependenceLowGrayLevelEmphasis (PTR) + 0.001304788 × wavelet.HHH_gldm_SmallDependenceLowGrayLevelEmphasis (PTR).

Intratumoral radiomics model

For intratumoral radiomics features, a total of 251 features were subjected to LASSO regression analysis with 10-fold cross validation for dimensionality reduction. The optimal regularization parameter λ was selected using the 1 standard error criterion, resulting in 9 nonzero features retained for model construction. The corresponding dimensionality reduction curve and LASSO coefficient convergence plots are presented in Figure 3C,3D. The intratumoral radiomic (ITR_RAD) model comprised of one morphological feature, one first order histogram feature, one texture feature, and six wavelet transformed features. The final radiomics score was calculated using the following formula: Rad-score = −0.66999884 + 0.28995121 × original.shapeSphericity + 0.58080224 × original.firstorderMedian + 0.06879913 × original.glcm_MCC − 0.03416258 × wavelet.LHL_firstorderMean − 0.03416258 × wavelet.HLL_firstorderMean + 0.32125690 × wavelet.HLL_glcm_MCC + 0.34045769 × wavelet.HLL_ngtdm_Strength + 0.12604603 × wavelet.LLL_glcm_DifferenceVariance + 0.08784636 × wavelet.LLL_ngtdm_Strength.

Peritumoral radiomic (PTR_RAD) model

For peritumoral radiomics features, a total of 153 features were analyzed. LASSO regression with 10-fold cross validation was applied for dimensionality reduction. Nine nonzero features were selected for construction of the peritumoral radiomics model using the 1 standard error criterion. The dimensionality reduction curve and LASSO coefficient convergence plots are presented in Figure 3E,3F. The PTR_RAD model comprised of two morphological features and seven wavelet transformed features. The final radiomics score was calculated as follows: Rad-score = −0.58831268 − 0.42602941 × original.shapeMaximum2DDiameterSlice + 0.04332238 × original.shapeSphericity − 0.03650785 × wavelet.LLH_glcm_Idn − 0.11170576 × wavelet.LHL_glcm_Imc1 + 0.29248361 × wavelet.LHH_ngtdm_Coarseness + 0.07486295 × wavelet.HLL_glcm_Imc2 + 0.14623326 × wavelet.HHL_glcm_Correlation − 0.12772086 × wavelet.LLL_glcm_Idn − 0.12772086 × wavelet.LLL_gldm_DependenceVariance.

Comparison of the three models

ROC curves for the MR_RAD are presented in Figure 3.2. In the training set, the AUC was 0.949 (95% CI: 0.910–0.988), with a sensitivity of 83.3% and an accuracy of 91.0%, as presented in Figure 4A. In the testing set, the AUC was 0.913 (95% CI: 0.812–1.000), with a sensitivity of 75.0% and an accuracy of 83.0%, as presented in Figure 4B. The concordance index for predicting lung nodule characteristics was 0.913 (95% CI: 0.885–0.942).

Figure 4 The ROC curves of three radiomics models in the training group (A) and testing group (B). AUC, area under the curve; ITR_RAD, intratumoral radiomic; MR_RAD, multiple region radiomic; PTR_RAD, peritumoral radiomic; ROC, receiver operating characteristic.

ROC curves for the ITR_RAD model are also presented in Figure 3.2. In the training set, the AUC was 0.952 (95% CI: 0.915–0.989), with a sensitivity of 83.3% and an accuracy of 90.1%, as presented in Figure 4A. In the testing set, the AUC was 0.927 (95% CI: 0.829–1.000), with a sensitivity of 81.3% and an accuracy of 87.2%, as presented in Figure 4B. The concordance index for predicting lung nodule characteristics was 0.927 (95% CI: 0.900–0.955). DeLong test comparisons between the MR_RAD and ITR_RAD models indicated that, although the AUC of the ITR_RAD model was slightly higher, no statistically significant difference in discriminative performance was observed in either the training set (P=0.33) or the testing set (P=0.20).

ROCs for the PTR_RAD model are presented in Figure 4. In the training set, the AUC was 0.907 (95% CI: 0.855–0.960), with a sensitivity of 78.6% and an accuracy of 82.0%, as presented in Figure 4A. In the testing set, the AUC was 0.742 (95% CI: 0.600–0.887), with a sensitivity of 56.3% and an accuracy of 70.2%, as presented in Figure 4B. DeLong test results demonstrated that the discriminative performance of the ITR_RAD model was significantly superior to that of the PTR_RAD model in the testing set (P=0.03), whereas the difference did not reach statistical significance in the training set (P=0.053).

From the comparisons of AUC values, sensitivity, specificity, positive predictive value, negative predictive value, and accuracy, as presented in Table 4, together with DeLong test results, the ITR_RAD model was selected for integration with clinical predictors, including molecular subtype and N stage, to construct a combined model. DeLong test comparisons indicated that the ITR_RAD model demonstrated improved discriminative performance relative to the clinical model in the training set (P<0.001), whereas no statistically significant improvement was observed in the testing set (P=0.08). The concordance index for predicting lung nodule characteristics in the combined clinical and ITR_RAD model was 0.742 (95% CI: 0.701–0.783), which was lower than the concordance indices observed for the MR_RAD and ITR_RAD models, indicating relatively lower discriminative ability of the combined model.

Integrated model construction

An integrated prediction model was developed by combining the radiomics score derived from the ITR_RAD with significant clinical predictors, including N stage and luminal subtype. In the training set, the integrated model achieved an area under the receiver operating characteristic curve of 0.959 (95% CI: 0.926–0.991), with a sensitivity of 81.0% and an accuracy of 90.1%. In the testing set, the AUC was 0.946 (95% CI: 0.850–1.000), with a sensitivity of 87.5% and an accuracy of 91.5%. When compared to the clinical prediction model, MR_RAD, ITR_RAD, and PTR_RAD, the integrated model demonstrated superior performance in terms of both AUC and accuracy, as presented in Table 4 and Figure 5. DeLong test results indicated that the integrated model provided a statistically significant improvement in discrimination compared with the clinical prediction model in both the training set (P<0.001) and the testing set (P=0.045). These findings indicate that the incorporation of radiomics features substantially enhanced predictive performance beyond the use of N stage and luminal subtype alone for classification of lung nodule pathology. The concordance index for the integrated model in predicting lung nodule characteristics was 0.945 (95% CI: 0.919–0.973). Calibration curves for the training and testing sets are presented in Figure 6 and demonstrated good agreement between predicted probabilities and observed outcomes in both cohorts. DCA was performed to compare the clinical use of the integrated model with that of the ITR_RAD model and the clinical model (Figure 7). Results from the testing set indicated that the integrated model yielded a higher net clinical benefit across a range of threshold probabilities, supporting its potential value for clinical application.

Figure 5 The ROC curves of the clinical model, the ITR_RAD model and the combined model in the training group (A) and testing group (B). AUC, area under the curve; ITR_RAD, intratumoral radiomic; ROC, receiver operating characteristic.

Figure 6 The calibration plot for the combined model in the training group (A) and testing group (B).

Figure 7 DCA for the clinical model, the ITR_RAD model and the combined model in the training group (A) and testing group (B). DCA, decision curve analysis; ITR_RAD, intratumoral radiomic.

Discussion

A clinical-radiomics integrated prediction model was developed to differentiate BCLM from PLC by incorporating clinicopathological characteristics and CT radiomic features. When compared to patients with PLC, patients with BCLM more frequently presented with higher T and N stages, increased HER2 positivity, higher Ki-67 expression, elevated serum CA153 levels, and a predominance of the HER2-positive molecular subtype. The radiomics models contributed to characterization of the pathological nature of pulmonary nodules in patients with breast cancer, with the intratumoral radiomic model demonstrating superior discriminatory performance compared with the peritumoral model. Integration of clinical predictors with radiomic features further improved predictive performance relative to either approach alone. This combined model supports earlier differentiation of pulmonary nodules, thereby facilitating timely and targeted therapeutic strategies.

Breast cancer has been the most frequently diagnosed malignancy among women in developed countries for more than two decades. In China, breast cancer has similarly surpassed lung cancer as the most prevalent cancer type among women, placing substantial social and economic burdens on the healthcare system (15). Mortality related to breast cancer is largely attributable to metastatic disease, particularly in triple-negative breast cancer, which exhibits a higher propensity for distant dissemination. Once metastasis occurs, the disease generally progresses to an incurable stage.

The lung is one of the most common sites of distant metastasis in breast cancer and is associated with a significant increase in mortality risk. Prognosis for patients with BCLM remains poor. A retrospective analysis of the Surveillance, Epidemiology, and End Results database reported a median survival of only 25 months among patients with breast cancer and isolated lung metastasis (16). Therefore, postoperative systemic therapy, regular surveillance, and early detection of lung metastasis followed by appropriate targeted treatment are critical for improving survival outcomes.

Advances in breast cancer diagnosis and treatment have prolonged survival, resulting in an increasing incidence of second primary malignancies. The annual incidence of second PLC among women with breast cancer has been reported to be 8.20 per 10,000 (17). However, no consensus definition exists for synchronous multiple primary cancers. Warren proposed in 1932 that tumors diagnosed within six months should be considered synchronous, whereas those diagnosed after 6 months should be regarded as metachronous (18). Subsequent studies have adopted alternative intervals, including 2 months, 3 months, or 1 year (19-21). Given evidence that the incidence of second PLC is highest within the first year following breast cancer diagnosis, a 1-year interval between breast and lung cancer diagnoses was used as the inclusion criterion in the present analysis (22).

Several studies have demonstrated prognostic differences between PLC and BCLM. Song et al. reported a higher three-year survival rate in patients with PLC compared to those with BCLM (23). Similarly, Zhao et al. observed significantly better disease-free survival among patients with PLC than among those with BCLM (7). Treatment for BCLM primarily relies on systemic therapies, including chemotherapy, endocrine therapy, and targeted therapy, whereas the role of pulmonary metastasectomy remains uncertain (24). Patients with BCLM are commonly diagnosed with stage IV breast cancer, whereas patients with PLC are more often diagnosed at earlier stages, and lesions are frequently solitary (25). Consequently, management of second PLC in breast cancer survivors generally emphasizes radical surgical resection. Accurate and early differentiation between BCLM and PLC is therefore essential to optimize treatment selection and improve clinical outcomes.

The American Joint Committee on Cancer staging system is widely used in breast cancer to guide diagnosis, prognosis, and treatment decisions. In the present cohort, a higher proportion of patients with PLC were classified as T1 (40.0% vs. 19.0%, P=0.004) compared with those with BCLM, whereas T3 and T4 stages (20.7% vs. 7.0%) were more frequent among patients with BCLM. Although T stage was associated with lung nodule pathology in univariate analysis, it was not retained as an independent predictor in multivariate modeling, indicating the need for further investigation into its role in lung metastasis. The N0 rate in PLC patients was significantly higher than in BCLM patients (80.0% vs. 41.4%, P<0.001), while BCLM patients had a higher proportion of N1 (51.7% vs. 19.0%) and N2 + N3 (6.9% vs. 1.0%) lymph node involvement. Higher N stage generally reflects increased tumor invasiveness and metastatic potential, which may contribute to lung dissemination. These findings are consistent with prior reports identifying N stage as an independent predictor of lung metastasis (6).

Univariate and multivariate logistic regression analyses indicated that luminal subtype was a significant factor associated with lung metastasis in breast cancer. The HER2-positive subtype was associated with an increased risk of lung metastasis, with statistical significance [odds ratio (OR) =4.27; 95% CI: 1.16–15.70, P=0.02]. In this cohort, 44.8% of patients with BCLM were HER2-positive, whereas only 16.0% of patients with PLC demonstrated HER2 positivity. An analysis of the Surveillance, Epidemiology, and End Results database reported that the HR+/HER2+ subtype exhibited a higher likelihood of lung metastasis compared with the HR+/HER2− subtype (26). HER2 signaling contributes to tumor cell proliferation and growth and facilitates stromal invasion and vascular entry, thereby promoting metastatic dissemination (27). Therefore, HER2 expression may provide important supplementary information for predicting lung metastasis. In contrast, no statistically significant difference in lung metastasis risk was observed between the HR+/HER2− and HR−/HER2− subtypes (OR =1.49; 95% CI: 0.47–4.71; P=0.49). Hagen et al. conducted a retrospective analysis of 3726 patients with metastatic breast cancer and reported higher risks of lung metastasis for basal-like subtypes (OR =2.5; 95% CI: 1.6–3.9) and non-basal triple-negative breast cancer (OR =2.1; 95% CI: 1.3–3.5) compared with the luminal A subtype (28). These findings differ from the present results, likely reflecting differences in patient populations and treatment eras. Hagen et al. assessed patients diagnosed between 1986 and 1992, whereas the present cohort included 25 HR−/HER2− patients, with 64% diagnosed after 2018. The widespread use of capecitabine and olaparib in the management of triple-negative breast cancer has significantly improved progression-free survival and overall survival, which may partially explain the lower representation of triple-negative cases in this study (29,30).

Differentiation of small pulmonary nodules, particularly those smaller than 1 cm in diameter, remains challenging using CT imaging alone. PLC nodules may present as ground-glass, solid, or part-solid lesions and may exhibit spiculation, lobulation, or vascular clustering. In contrast, BCLM typically manifests as solid nodules with smooth margins and relatively uniform density, with cavitation being uncommon (31). In clinical practice, definitive differentiation of isolated small lung nodules often requires CT-guided biopsy or surgical resection for pathological confirmation. Chang et al. reported that only 30% of isolated lung nodules detected in patients with breast cancer represented BCLM, whereas 70% were confirmed to be PLC (32). Misclassification of PLC as BCLM may delay radical surgical management, while misclassification of BCLM as PLC may lead to under recognition of advanced breast cancer and inappropriate adjustment of systemic therapies, including chemotherapy, endocrine therapy, and targeted therapy. Therefore, early identification and accurate characterization of isolated pulmonary nodules are critical for supporting clinical decision making and enabling timely, appropriate therapeutic approaches.

Radiomics enables conversion of medical images into high-throughput quantitative features (33). Through the application of machine learning methods for feature selection and dimensionality reduction, radiomics aims to develop robust models for clinical tasks such as diagnosis, prognosis, and evaluation of treatment response. As a non-invasive approach, radiomic features reflect multiple biological characteristics of tumors, including intratumoral and intertumoral heterogeneity, features of the tumor microenvironment, and patterns associated with infiltrating cells (34). The application of radiomics and artificial intelligence in breast cancer has expanded to include prediction of molecular subtypes, axillary lymph node metastasis, and response to neoadjuvant chemotherapy (35-37). Radiomics has also been increasingly applied in lung cancer diagnosis and clinical decision making (10,38). However, limited research has focused on using radiomics to differentiate the pathological nature of lung nodules in patients with breast cancer. Prior studies attempting to distinguish lung metastases originating from colorectal, breast, and renal cancers reported that differences were primarily driven by local gray-level variations between adjacent pixels, with GLCM features being particularly informative, consistent with the present findings (11).

In this analysis, the radiomic features with the largest coefficients in the intratumoral radiomic model included original.firstorderMedian, wavelet.HLL_glcm_MCC, and wavelet.HLL_ngtdm_Strength. These features represent average gray-level intensity within the ROI, texture roughness based on GLCM metrics, and strength derived from neighboring gray tone difference matrix features, respectively. Higher radiomics scores, reflecting increased textural complexity and more pronounced voxel-level variation, were associated with a higher likelihood of metastatic nodules, consistent with prior findings distinguishing primary from metastatic lung tumors (39).

Currently, few studies have applied radiomics to differentiate PLC from BCLM. Li et al. conducted a retrospective analysis that included 60 patients with BCLM and 40 patients with PLC and developed three region-based radiomics models using intratumoral, peritumoral, and mixed regions (40). The reported AUC values were 0.798 (95% CI: 0.706–0.872) for the intratumoral model, 0.814 (95% CI: 0.724–0.885) for the peritumoral model, and 0.856 (95% CI: 0.772–0.918) for the multi-region model. Our finding that the intratumoral radiomics (ITR) model outperformed the PTR and mixed-region (MR) models appears to contrast with a previous study (40), which reported the lowest performance for the ITR model. Several factors may explain this discrepancy. First, the clinical task differed substantially: Li et al. (40) distinguished benign from malignant pulmonary nodules, whereas our study differentiated two malignant entities—BCLM and PLC. In benign versus malignant discrimination, peritumoral features such as spiculation and ground-glass opacity are highly sensitive and specific, whereas distinguishing between two cancer types may rely more on intratumoral heterogeneity. Second, segmentation protocols varied; Li et al. (40) used fixed radial expansions, which may include irrelevant structures (e.g., vessels, bronchi), whereas we employed shape-adaptive PTRs with exclusion of large vessels, reducing noise. Third, our cohort was histopathologically homogeneous (all breast cancer patients with confirmed lung lesions), while Reference 40 included a broader spectrum of pulmonary nodules. Finally, differences in radiomic software, preprocessing steps, and sample size may also contribute to the observed discrepancy. These comparisons highlight the importance of task-specific and cohort-specific considerations when interpreting radiomic model performance.

The objective of the present analysis was to develop a clinical-radiomics integrated model to differentiate PLC from BCLM and thereby improve diagnostic accuracy and clinical decision making. In the testing cohort, the integrated model achieved an AUC of 0.946 (95% CI: 0.850–1.000), exceeding the performance of the intratumoral radiomic model (AUC =0.927; 95% CI: 0.829–1.000) and the clinical model (AUC =0.760; 95% CI: 0.615–0.905). This performance also surpassed that reported in prior radiomics studies (41). The integrated model demonstrated a sensitivity of 87.5%, specificity of 93.5%, and an accuracy of 91.5% in the testing cohort, indicating superior overall performance compared with individual models. The concordance index of 0.945 (95% CI: 0.919–0.973) further indicated strong agreement between predicted and observed pathological outcomes. DCA confirmed that the integrated model provided a higher net clinical benefit than individual models, supporting its potential applicability in clinical practice.

In the present study, the combined model misclassified 8 out of 158 patients (5.1%), including five patients with pathologically confirmed PLC who were incorrectly predicted as having BCLM and three patients with BCLM who were incorrectly predicted as having PLC. A detailed review of these misclassified cases revealed several recurrent patterns. For PLC cases misclassified as BCLM, the lung lesions frequently appeared as solitary, peripheral nodules with well-defined borders and lacked typical malignant features such as spiculation or lobulation. Notably, these patients often had a history of HER2-positive breast cancer, a subtype associated with increased lung tropism, which may have biased the model toward overestimating the probability of metastasis. Conversely, BCLM cases misclassified as PLC commonly presented with multiple lung lesions showing irregular margins, spiculation, or cavitation—radiological features more characteristic of primary lung malignancy. Clinically, these patients had a longer interval between the initial breast cancer diagnosis and the detection of lung lesions (mean >5 years) and lower serum levels of tumor markers such as CA15-3, making differentiation difficult even for experienced radiologists. In addition, small lesion size (<1 cm) and the presence of pre-existing benign lung conditions (e.g., emphysema or pulmonary fibrosis) were associated with higher misclassification rates in both directions, likely due to reduced radiomic signal and feature interference, respectively. These observations highlight the inherent challenges in distinguishing BCLM from PLC using CT-based radiomics alone, particularly in atypical presentations or in the presence of confounders. Future efforts should focus on incorporating dynamic contrast-enhanced CT or positron emission tomography (PET)/CT features, as well as expanding sample sizes to enable subgroup analyses that may further refine model performance.

Several limitations should be acknowledged. The sample size was relatively limited, which may have introduced bias in model development and evaluation. Larger cohorts are required to validate these findings. Patients with benign lung nodules confirmed by biopsy were not included, limiting the assessment of model performance across a broader clinical spectrum. In addition, the retrospective single-center design may restrict generalizability. Manual delineation of ROI using 3D Slicer may have introduced observer-related variability and required substantial time, potentially limiting scalability. Future studies should incorporate automated segmentation approaches to improve efficiency and reproducibility. Furthermore, some clinical variables, including additional pathological characteristics and nodule location, were not incorporated into the model.

Conclusions

This study provides a clinically relevant integrated model for distinguishing BCLM from PLC and may offer meaningful support for diagnostic assessment and treatment planning in patients with breast cancer.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0600/rc

Data Sharing Statement: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0600/dss

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0600/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0600/coif). The authors have no conflicts of interest to declare.

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 study was approved by the Institutional Review Board of the Affiliated Union Hospital of Fujian Medical University (No. 2025KY029). Informed consent was taken from all the patients (or a statement that it was not required and why).

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Siegel RL, Giaquinto AN, Jemal A. Cancer statistics, 2024. CA Cancer J Clin 2024;74:12-49. [Crossref] [PubMed]
2. Huang QY, Liu GF, Qian XL, et al. Long Non-Coding RNA: Dual Effects on Breast Cancer Metastasis and Clinical Applications. Cancers (Basel) 2019;11:1802. [Crossref] [PubMed]
3. Hayat MJ, Howlader N, Reichman ME, et al. Cancer statistics, trends, and multiple primary cancer analyses from the Surveillance, Epidemiology, and End Results (SEER) Program. Oncologist 2007;12:20-37. [Crossref] [PubMed]
4. Soh J, Komoike Y, Mitsudomi T. Surgical therapy for pulmonary metastasis of breast cancer. Transl Cancer Res 2020;9:5044-52. [Crossref] [PubMed]
5. Miyazaki K, Satoh H, Watanabe H, et al. Sole metastatic pulmonary nodules from breast cancer simulating primary lung adenocarcinoma: Two case reports. Mol Clin Oncol 2017;6:425-7. [Crossref] [PubMed]
6. He L, Wang X, Liu X, et al. Analysis of Clinical Characteristics, Treatment, and Prognostic Factors of 106 Breast Cancer Patients With Solitary Pulmonary Nodules. Front Surg 2022;9:843913. [Crossref] [PubMed]
7. Zhao W, Song C, Zhu S, et al. Prognostic implication of isolated pulmonary nodules in patients with a history of breast cancer. Journal of Cardiothoracic Surgery 2022;17:160.
8. He KW, Wei W, Liu ZY, et al. Clinicopathological features of second primary lung cancer and pulmonary metastasis in patients with breast cancer. Zhonghua Zhong Liu Za Zhi 2018;40:201-5. [Crossref] [PubMed]
9. Mayerhoefer ME, Materka A, Langs G, et al. Introduction to Radiomics. J Nucl Med 2020;61:488-95. [Crossref] [PubMed]
10. Zhong F, Liu Z, An W, et al. Radiomics Study for Discriminating Second Primary Lung Cancers From Pulmonary Metastases in Pulmonary Solid Lesions. Front Oncol 2021;11:801213. [Crossref] [PubMed]
11. Shang H, Li J, Jiao T, et al. Differentiation of Lung Metastases Originated From Different Primary Tumors Using Radiomics Features Based on CT Imaging. Acad Radiol 2023;30:40-6. [Crossref] [PubMed]
12. Miao L, Ma ST, Jiang X, et al. Prediction of the therapeutic efficacy of epirubicin combined with ifosfamide in patients with lung metastases from soft tissue sarcoma based on contrast-enhanced CT radiomics features. BMC Med Imaging 2022;22:131. [Crossref] [PubMed]
13. Yu Y, Tan J, Yang Y, et al. The Differential Diagnostic Value of Radiomics Signatures Between Single-Nodule Pulmonary Metastases and Second Primary Lung Cancer in Patients with Colorectal Cancer. Technol Cancer Res Treat 2023;22:15330338231175735. [Crossref] [PubMed]
14. Li J, Liu L, Feng Z, et al. Tumor markers CA15-3, CA125, CEA and breast cancer survival by molecular subtype: a cohort study. Breast Cancer 2020;27:621-30. [Crossref] [PubMed]
15. Karakiewicz PI, Benayoun S, Kattan MW, et al. Development and validation of a nomogram predicting the outcome of prostate biopsy based on patient age, digital rectal examination and serum prostate specific antigen. J Urol 2005;173:1930-4. [Crossref] [PubMed]
16. Zheng RS, Zhang SW, Sun KX, et al. Cancer statistics in China, 2016. Zhonghua Zhong Liu Za Zhi 2023;45:212-20. [Crossref] [PubMed]
17. Xiao W, Zheng S, Liu P, et al. Risk factors and survival outcomes in patients with breast cancer and lung metastasis: a population-based study. Cancer Med 2018;7:922-30. [Crossref] [PubMed]
18. Lin FW, Yeh MH, Lin CL, et al. Association between Breast Cancer and Second Primary Lung Cancer among the Female Population in Taiwan: A Nationwide Population-Based Cohort Study. Cancers (Basel) 2022;14:2977. [Crossref] [PubMed]
19. Zhang WX, Ji Y, Zhai J, et al. Clinical and pathological features of breast and lung double primary cancers. Mod Oncol Med 2021;29:427-31.
20. Grossman D, Farnham JM, Hyngstrom J, et al. Similar survival of patients with multiple versus single primary melanomas based on Utah Surveillance, Epidemiology, and End Results data (1973-2011). J Am Acad Dermatol 2018;79:238-44. [Crossref] [PubMed]
21. Pomerantz H, Huang D, Weinstock MA. Risk of subsequent melanoma after melanoma in situ and invasive melanoma: a population-based study from 1973 to 2011. J Am Acad Dermatol 2015;72:794-800. [Crossref] [PubMed]
22. Moseley HS, Giuliano AE, Storm FK 3rd, et al. Multiple primary melanoma. Cancer 1979;43:939-44. [Crossref] [PubMed]
23. Wang R, Yin Z, Liu L, et al. Second Primary Lung Cancer After Breast Cancer: A Population-Based Study of 6,269 Women. Front Oncol 2018;8:427. [Crossref] [PubMed]
24. Song Z, Ye T, Ma L, et al. Surgical Outcomes of Isolated Malignant Pulmonary Nodules in Patients with a History of Breast Cancer. Ann Surg Oncol 2017;24:3748-53. [Crossref] [PubMed]
25. Tanaka F, Li M, Hanaoka N, et al. Surgery for pulmonary nodules in breast cancer patients. Ann Thorac Surg 2005;79:1711-5.
26. He YY, Chen X, Zhang W, et al. Epidemiological and clinical pathological features of secondary lung cancer in breast cancer patients. Chin J Surg Oncol 2018;10:291-5.
27. Wu Q, Li J, Zhu S, et al. Breast cancer subtypes predict the preferential site of distant metastases: a SEER based study. Oncotarget 2017;8:27990-6. [Crossref] [PubMed]
28. Kedrin D, Wyckoff J, Boimel PJ, et al. ERBB1 and ERBB2 have distinct functions in tumor cell invasion and intravasation. Clin Cancer Res 2009;15:3733-9. [Crossref] [PubMed]
29. Kennecke H, Yerushalmi R, Woods R, et al. Metastatic behavior of breast cancer subtypes. J Clin Oncol 2010;28:3271-7. [Crossref] [PubMed]
30. Wu JN, Li L, Huang XW, et al. Clinical observation of capecitabine maintenance therapy in recurrent and metastatic triple-negative breast cancer. Anhui Med J 2016;20:1181-4.
31. Li B, Zhang GF. Advances in diagnosis and treatment of late-stage triple-negative breast cancer and future directions. Henan Med Res 2023;32:3636-44.
32. Shen SJ, Han ZJ, Yu SN, et al. Breast cancer combined with malignant lung nodules: Breast cancer lung metastasis or breast and lung double primary cancer? Peking Union Med Coll J 2020;11:109-14.
33. Chang EY, Johnson W, Karamlou K, et al. The evaluation and treatment implications of isolated pulmonary nodules in patients with a recent history of breast cancer. Am J Surg 2006;191:641-5. [Crossref] [PubMed]
34. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278:563-77. [Crossref] [PubMed]
35. Pesapane F, Rotili A, Agazzi GM, et al. Recent Radiomics Advancements in Breast Cancer: Lessons and Pitfalls for the Next Future. Curr Oncol 2021;28:2351-72. [Crossref] [PubMed]
36. Zhang Q, Peng Y, Liu W, et al. Radiomics Based on Multimodal MRI for the Differential Diagnosis of Benign and Malignant Breast Lesions. J Magn Reson Imaging 2020;52:596-607. [Crossref] [PubMed]
37. Yang J, Wang T, Yang L, et al. Preoperative Prediction of Axillary Lymph Node Metastasis in Breast Cancer Using Mammography-Based Radiomics Method. Sci Rep 2019;9:4429. [Crossref] [PubMed]
38. Braman NM, Etesami M, Prasanna P, et al. Intratumoral and peritumoral radiomics for the pretreatment prediction of pathological complete response to neoadjuvant chemotherapy based on breast DCE-MRI. Breast Cancer Research 2017;19:57.
39. Liu S, Liu S, Zhang C, et al. Exploratory Study of a CT Radiomics Model for the Classification of Small Cell Lung Cancer and Non-small-Cell Lung Cancer. Front Oncol 2020;10:1268. [Crossref] [PubMed]
40. He XX. Radiomics studies in pulmonary metastatic tumors. Guangzhou Med Univ; 2023.
41. Li L, Zhou X, Cui W, et al. Combining radiomics and deep learning features of intra-tumoral and peri-tumoral regions for the classification of breast cancer lung metastasis and primary lung cancer with low-dose CT. J Cancer Res Clin Oncol 2023;149:15469-78. [Crossref] [PubMed]