Habitat-based ultrasound radiomics for preoperative prediction of axillary lymph node metastasis in breast cancer: a two-center study Authors: Zimeng Fang, Xiaoyu Yan, Yunyun Zhan, Yu Bi, Xiabi Wu, Mingdi Fang, Yufan Zhang, Jiajia Wang, Mei Peng Published: 2026 ABSTRACT BackgroundAccurate preoperative prediction of axillary lymph node metastasis (ALNM) remains challenging. This study investigates whether habitat-based ultrasound radiomics of suspicious axillary lymph nodes can predict preoperative ALNM in patients with breast cancer.MethodsThis retrospective two-center study included 416 breast cancer patients with suspicious axillary lymph nodes and pathologically confirmed nodal status. Patients from the Second Affiliated Hospital of Anhui Medical University were randomly assigned to the training and internal validation cohorts at a ratio of 7:3, whereas patients from Taihe County People’s Hospital constituted the external test cohort. Suspicious lymph nodes were manually segmented. Pixel-level features were clustered using the K-means algorithm to divide each node into four habitat subregions. Radiomics features were extracted from the intranodal region of interest and each habitat subregion to construct the intranodal radiomics, individual habitat, and habitat radiomics models. A combined nomogram integrating the optimal radiomics signature with independent clinical predictors was developed.ResultsCortical thickness and lymph node vascular pattern were independent predictors of ALNM. The habitat radiomics model outperformed the intranodal radiomics and individual habitat models, with areas under the receiver operating characteristic curve (AUCs) of 0.859, 0.809, and 0.791 in the training, internal validation, and external test cohorts, respectively. The combined nomogram achieved the highest AUCs of 0.866, 0.819, and 0.806 in the corresponding cohorts, and its clinical utility was confirmed by decision curve analysis.ConclusionsHabitat-based ultrasound radiomics improved ALNM prediction by capturing intranodal heterogeneity. The nomogram may provide a practical noninvasive tool for individualized preoperative assessment of suspicious axillary lymph nodes. Highlight box Key findings • This two-center study developed and externally validated a habitat-based ultrasound radiomics model using two-dimensional ultrasound images of suspicious axillary lymph nodes to predict axillary lymph node metastasis in breast cancer. The combined nomogram integrating the habitat radiomics signature, cortical thickness, and lymph node vascular pattern achieved good predictive performance, with area under the curve values of 0.866, 0.819, and 0.806 in the training, internal validation, and external test cohorts, respectively. What is known and what is new? • Previous studies have shown that ultrasound-based radiomics can assist in preoperative assessment of axillary lymph node status. However, most models focus on the primary tumor or extract global features from the entire lymph node region of interest, which may overlook intranodal spatial heterogeneity. • This study introduced habitat analysis into suspicious axillary lymph nodes on ultrasound. By dividing each node into four subregions, the model captured regional heterogeneity and provided more individualized prediction than conventional intranodal radiomics. What is the implication, and what should change now? • The nomogram provides clinicians with a noninvasive tool for preoperative risk stratification of suspicious axillary lymph nodes. It may support individualized axillary management, but prospective multicenter validation and histopathologic correlation are needed before routine clinical implementation. Introduction Breast cancer remains the most commonly diagnosed malignancy among women worldwide and a leading cause of cancer-related death. In 2020, an estimated 2.3 million new cases of female breast cancer and 685,000 related deaths were reported globally (1). In breast cancer patients, the occurrence of axillary lymph node (ALN) metastasis is closely related to the overall survival. Compared with patients without axillary lymph node metastasis (ALNM), those with ALNM have a significantly lower 5-year survival rate (2). The presence of ALNM affects the determination of clinical surgical plans, the potential demand for adjuvant therapy, and the ultimate prognosis of patients (3,4). Contemporary axillary assessment integrates clinical examination and imaging. Sonographically suspicious lymph nodes may undergo ultrasound-guided needle biopsy, whereas sentinel lymph node biopsy (SLNB) is used for surgical staging in appropriately selected clinically node-negative patients (5). The SOUND and INSEMA trials support the omission of SLNB in carefully selected clinically node-negative patients (6,7), while SENOMAC indicates that limited sentinel lymph node metastasis does not necessarily require completion of axillary lymph node dissection (8). Collectively, these findings underscore the importance of accurate patient selection and individualized risk assessment. Accordingly, noninvasive quantification of nodal metastatic risk may help address this need by complementing conventional ultrasonography and providing additional information for subsequent evaluation and treatment planning. Radiomics, by high-throughput extraction of quantitative imaging features, can reflect the intensity distribution, texture patterns and underlying tumor heterogeneity (9). By integrating these features with statistical or machine learning algorithms, radiomics models provide an objective framework for clinical decision-making (10,11). In recent years, ultrasound-based radiomics and deep learning approaches have been increasingly applied to the preoperative assessment of ALNM in breast cancer, demonstrating consistent predictive value (12-14). Nevertheless, most existing studies focus on the primary tumor or extract global features from whole region of interest (ROI) of lymph nodes (15,16), assuming spatial uniformity within the lesion. However, suspicious axillary nodes frequently exhibit complex internal changes (17). Relying solely on whole ROI analysis leads to a loss of spatial specificity, which can obscure localized metastatic details. These limitations highlight the necessity for advanced imaging techniques to characterize intranodal heterogeneity. Habitat analysis, as an emerging heterogeneity assessment strategy, can partition the tumor into multiple subregions with similar imaging phenotypes based on voxel-level or pixel-level features, and extract features from each subregion separately to characterize the internal spatial heterogeneity (18). This approach has demonstrated utility in tumor classification and treatment response assessment across various malignancies, including lung, ovarian, and breast cancers (19-21). Similarly, habitat analysis based on magnetic resonance imaging (MRI) has successfully quantified tumor heterogeneity and predicted responses to neoadjuvant chemotherapy in breast cancer (22). A recent bi-center study demonstrated that habitat analysis of axillary lymph node ultrasound images could capture intranodal heterogeneity and achieve favorable performance in external validation (23). Nevertheless, evidence remains limited regarding the integration of habitat-derived signatures with conventional ultrasound features in externally validated nomograms. This study aimed to apply habitat analysis to suspicious axillary lymph nodes on two-dimensional ultrasound images and develop predictive models. Furthermore, we integrated clinical information to build a combined model and constructed a nomogram for individualized preoperative prediction of ALNM. We present this article in accordance with the TRIPOD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0363/rc). Methods Patients The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This retrospective multicenter study was approved by the Institutional Review Board (IRB) and Ethics Committee of The Second Affiliated Hospital of Anhui Medical University (Center 1). Taihe County People’s Hospital (Center 2) was informed of and agreed to participate in this study. The requirement for informed consent was waived due to the retrospective nature of the study. The inclusion criteria were as follows: (I) patients with newly diagnosed primary breast cancer who had not received any treatment prior to ultrasound (US) examination, including neoadjuvant chemotherapy, radiotherapy, endocrine therapy, or targeted therapy; (II) the presence of at least one suspicious axillary lymph node on preoperative US, with the target lymph node selected based on overall sonographic assessment; when multiple suspicious lymph nodes were present, the largest suspicious lymph node was selected; (III) The target lymph node underwent ultrasound guided biopsy for pathological confirmation, and complete, high-quality 2D ultrasound images of the same node taken before biopsy were available for habitat analysis. The exclusion criteria included: (I) absence of ultrasound-guided pathological confirmation of the target ALN; (II) received neoadjuvant chemotherapy, radiotherapy, or other therapies prior to US examination; (III) patients with incomplete clinical or pathological data; (IV) lymph nodes not fully included within the imaging field for habitat analysis. Finally, a total of 416 patients were enrolled. The cases from Center 1 were randomly divided into a training set and an internal validation set at a ratio of 7:3 (training set, n=215; validation set, n=93), whereas Center 2 served as the external test set (Figure 1). Flow diagram of patient enrollment. Center 1: The Second Affiliated Hospital of Anhui Medical University; Center 2: Taihe County People’s Hospital. ALNs, axillary lymph nodes. Clinical information The clinical information was retrospectively collected from the institutional electronic medical record systems of both centers. Variables included age at diagnosis, family history of breast cancer, tumor laterality (left or right breast), and quadrants (upper outer quadrant, non-upper outer quadrant). Molecular subtypes were classified as Luminal A, Luminal B, HER2 positive, and triple-negative breast cancer (TNBC). Ki-67 status was categorized as <14% or ≥14% (24). Image acquisition and analysis A comprehensive preoperative US examination of the breast and axillary regions was performed using high-resolution diagnostic systems, including the Samsung RS85A, Samsung RS80A, Samsung HERAW9 and GE LOGIQE10s equipped with high-frequency linear array transducers. All examinations were conducted according to standard breast imaging protocols at the two centers. Two-dimensional US images of the primary breast tumors and target ALNs obtained before biopsy were retrospectively retrieved from the Picture Archiving and Communication System (PACS). For each patient, the target ALN was matched to the ultrasound-guided biopsy record to confirm that the analyzed node was the pathologically sampled node. From the breast ultrasound images, the following conventional ultrasound features were recorded: tumor size, orientation (parallel or nonparallel), and calcification status (non-calcification, macrocalcifications, or microcalcifications). For target ALN, cortical thickness (<3 or ≥3 mm) and lymph node vascular pattern (hilar vs. mixed/peripheral) were recorded. These clinical and ultrasound features were included in the subsequent analyses and evaluated for clinical model development. Image pre-processing and ROI segmentation To reduce non-biological variability caused by differences in ultrasound devices and acquisition parameters, all images were resampled to a uniform pixel spacing of 1×1 mm before feature extraction. The US images were loaded into the ITK-SNAP 3.8.0 software (http://www.itksnap.org) for manual ROI segmentation. Two senior physicians with over 10 years of experience in breast ultrasound segmented the ROI of the target ALN. All the US experts were blinded to the clinical and pathological results of the study population. Sixty ultrasound images were randomly selected to calculate the intraclass correlation coefficient (ICC) to assess the reliability and reproducibility. An ICC ≥0.85 indicated good stability of the ROI, making it suitable for ultrasound radiomics feature extraction. Habitat generation A total of 19 quantitative features were extracted from each pixel within the segmented ROI to characterize local intensity and textural heterogeneity. To ensure consistency of habitat definition across patients, pixel clustering was conducted at the population level. Cluster numbers ranging from 2 to 9 were evaluated in the training cohort. The optimal number of habitat subregions was determined using the Calinski-Harabasz (CH) index, with a higher CH value indicating better within-cluster compactness and between-cluster separation (25). The cluster number with the highest CH value was selected for subsequent analysis. Feature extraction and selection Radiomics features were extracted from the segmented lymph node ROIs using the PyRadiomics package implemented in Python, following the feature definitions proposed by the Imaging Biomarker Standardization Initiative (IBSI). A total of 107 handcrafted radiomics features were extracted from each ROI. All extracted radiomics features were used to characterize the intensity distribution and spatial heterogeneity of the lymph node regions. Prior to feature selection, all radiomics features were standardized using Z-score normalization. All feature-selection procedures were performed exclusively within the training cohort. Univariate analysis with independent-sample t-tests was performed, and features with P<0.05 were retained. Pearson correlation analysis was subsequently used to remove redundant features, with one of each pair of highly correlated features (correlation coefficient >0.9) eliminated using a greedy recursive deletion strategy. The remaining features were further reduced using the minimum redundancy maximum relevance (mRMR) algorithm and least absolute shrinkage and selection operator (LASSO) logistic regression with ten-fold cross-validation to identify the optimal predictive feature subset. Model construction and evaluation Radiomics signatures were derived from the selected features extracted from the entire lymph node ROI and habitat subregions to establish imaging models. Univariate and multivariable logistic regression analyses were conducted on clinical and ultrasound characteristics to identify independent predictors of ALN metastasis and to develop the clinical model. Variables with a univariate P<0.05 were entered into multivariable logistic regression. The independent clinical predictors identified in the multivariable analysis were then combined with the Habitat Rad-score to construct a nomogram. The events per variable (EPV) ratio was calculated for the training cohort as the number of ALNM events divided by the number of predictors retained in the final combined model. Model performance was evaluated using receiver operating characteristic (ROC) curves, with the area under the ROC curve (AUC) quantifying predictive accuracy for the training, internal validation, and external test cohorts. A 95% confidence interval (CI) for each AUC was calculated using 1,000 bootstrap samples. Calibration curves were utilized to assess concordance between predicted and observed ALN metastasis outcomes. Decision curve analysis (DCA) was employed to determine the net benefit of each model across various threshold probabilities, thereby assessing the clinical utility of each model. The flowchart of the materials and methods in this study is shown in Figure 2. Overall workflow of habitat-based ultrasound radiomics analysis. Center 1: The Second Affiliated Hospital of Anhui Medical University; Center 2: Taihe County People’s Hospital. AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; MLP, multilayer perceptron; ROC, receiver operating characteristic; ROI, region of interest. Statistical analysis SPSS 26.0 software was used for statistical analysis. Measurements with the normal distribution were expressed as mean ± SD and analyzed by an independent sample t-test; measurements that do not fit the normal distribution were expressed as M (Q1, Q3) and analyzed by the Mann-Whitney U test. Count data were expressed as n (%) and analyzed by chi-square test. Variables with a P value <0.05 in univariate regression analysis were included in multivariable analysis. Variables with a P value <0.05 in multivariable analysis were considered independent predictors of ALN metastasis. Pairwise AUC comparisons were performed using the DeLong test. Brier scores were calculated to assess overall prediction accuracy, with lower values indicating better performance. A two-sided P<0.05 was considered statistically significant. Results Baseline characteristics A total of 416 patients were included. The training cohort comprised 215 patients, and the internal validation cohort comprised 93 patients from Center 1 (124 negative, 184 positive in total). The external test cohort included 108 patients from Center 2 (49 negative, 59 positive). The baseline clinical and ultrasound characteristics of the patients are shown in Table 1. No significant differences were observed among the training, validation, and external test cohorts in age, tumor size, family history, laterality, tumor orientation, tumor quadrant, calcification, cortical thickness, lymph node vascular pattern, molecular type, or Ki-67 status (all P>0.05). Table 1Baseline clinical and US characteristics of patients in training, validation and external test setsCharacteristicsTraining setValidation setExternal test setP valueAge, years53.01±10.8252.76±12.4153.25±11.780.96Tumor size, mm30.69±13.3230.04±15.3630.39±16.700.94Family history0.96 No210 (97.67)91 (97.85)106 (98.15) Yes5 (2.33)2 (2.15)2 (1.85)Laterality0.28 Right98 (45.58)39 (41.94)57 (52.78) Left117 (54.42)54 (58.06)51 (47.22)Orientation0.17 Parallel135 (62.79)55 (59.14)56 (51.85) Nonparallel80 (37.21)38 (40.86)52 (48.15)Quadrants0.42 Upper outer quadrant113 (52.56)50 (53.76)65 (60.19) Non-upper outer quadrant102 (47.44)43 (46.24)43 (39.81)Calcification0.31 Non-calcification85 (39.53)38 (40.86)35 (32.41) Macrocalcifications11 (5.12)3 (3.23)2 (1.85) Microcalcifications119 (55.35)52 (55.91)71 (65.74)Cortical thickness0.10 <3 mm39 (18.14)26 (27.96)28 (25.93) ≥3 mm176 (81.86)67 (72.04)80 (74.07)Lymph node vascular pattern0.24 Hilar40 (18.60)23 (24.73)23 (21.30) Mixed/peripheral175 (81.40)70 (75.27)85 (78.70)Molecular type0.37 Luminal A3 (1.40)2 (2.15)2 (1.85) Luminal B163 (75.81)65 (69.89)84 (77.78) HER2 positive31 (14.42)22 (23.66)17 (15.74) TNBC18 (8.37)4 (4.30)5 (4.63)Ki-670.47 <14%19 (8.84)11 (11.83)14 (12.96) ≥14%196 (91.16)82 (88.17)94 (87.04)Data are presented as mean ± standard deviation or n (%). HER2, human epidermal growth factor receptor 2; TNBC, triple-negative breast cancer; US, ultrasound. Baseline clinical and US characteristics of patients in training, validation and external test sets Data are presented as mean ± standard deviation or n (%). HER2, human epidermal growth factor receptor 2; TNBC, triple-negative breast cancer; US, ultrasound. Establishment of clinical models Univariate analysis demonstrated that several clinical and ultrasound features, including age, tumor size, molecular type, calcification, orientation, tumor location, quadrants, cortical thickness, lymph node vascular pattern, and Ki-67, were significantly associated with ALN metastasis in the training cohort (P<0.05). Subsequent multivariable logistic regression analysis revealed that cortical thickness and lymph node vascular pattern remained independent predictors for ALN metastasis (P<0.05), as shown in Table 2. The clinical model achieved AUCs of 0.681, 0.662, and 0.646 in the three cohorts, respectively. Table 2Univariate logistic regression and multivariate logistic regression analysisCharacteristicsUnivariate analysisMultivariate analysisOR (95% CI)P valueOR (95% CI)P valueFamily history0.667 (0.149–2.992)0.66––Age1.009 (1.004–1.013)0.0010.979 (0.961–0.998)0.06Tumor size1.014 (1.007–1.021)0.0010.992 (0.973–1.011)0.50Molecular type1.215 (1.102–1.339)0.0010.86 (0.611–1.21)0.47Calcification1.387 (1.185–1.623)0.0011.078 (0.825–1.409)0.64Orientation1.581 (1.083–2.305)0.0460.919 (0.535–1.582)0.80Laterality1.721 (1.255–2.358)0.0051.136 (0.685–1.885)0.68Quadrants2.0 (1.416–2.826)0.0011.21 (0.728–2.008)0.54Cortical thickness2.2 (1.684–2.875)<0.0013.436 (1.763–6.693)0.002Lymph node vascular pattern2.125 (1.627–2.773)<0.0012.555 (1.379–4.73)0.01Ki-671.722 (1.35–2.197)<0.0011.349 (0.573–3.18)0.57CI, confidence interval; OR, odds ratio. Univariate logistic regression and multivariate logistic regression analysis CI, confidence interval; OR, odds ratio. Subregion cluster and feature selection Based on the CH score calculated in the training cohort, the optimal number of clusters for the K-means algorithm was determined to be 4, as shown in Figure 3. Consequently, the intranodal ROI of suspicious axillary lymph nodes was divided into four habitat subregions (H1-H4), as illustrated in Figure 4, for subsequent feature extraction and model construction. Subsequently, the fitted K-means model was applied to the internal validation and the external test set. Calinski-Harabasz score plot for determining the optimal number of clusters. Representative ultrasound images and habitat partition maps of pathologically positive lymph nodes. In the training cohort, 107 handcrafted radiomics features were extracted from the whole ROI and from each of the four habitat subregions (H1-H4). The features derived from the four subregions were aggregated to form the habitat feature pool, which consisted of 428 radiomics features in total (107×4). The radiomics features ultimately retained in the Habitat model and their corresponding coefficients are shown in Figure 5. Selected handcrafted radiomic features and corresponding coefficients for the habitat model. Comparison of different radiomics models In the training cohort, the habitat radiomics model showed the highest discriminative performance (AUC, 0.859; 95% CI: 0.8108–0.9081), outperforming the intranodal radiomics model (AUC, 0.761; 95% CI: 0.6956–0.8271) and the individual habitat model (AUC range, 0.787–0.851). This pattern remained consistent in both the validation cohort (habitat AUC, 0.809; 95% CI: 0.7179–0.8994) and the external test cohort (habitat AUC, 0.791; 95% CI: 0.7769–0.8051), as shown in Table 3. Table 3Performance of the clinical, intranodal radiomics, habitat radiomics, and combined nomogram models in the training, internal validation, and external test cohortsCohort/modelAUC (95% CI)SensitivitySpecificityPPVNPVAccuracyTraining set (n=215) Clinic0.681 (0.6173–0.7443)0.8410.5060.730.6670.712 Intranodal0.761 (0.6956–0.8271)0.8330.5660.7530.6810.73 Habitat0.859 (0.8108–0.9081)0.6970.8670.8930.6430.763 Habitat H10.787 (0.7250–0.8495)0.8260.6270.7790.6930.749 Habitat H20.804 (0.7461–0.8625)0.6740.7710.8240.5980.712 Habitat H30.815 (0.7545–0.8746)0.9020.590.7780.790.781 Habitat H40.851 (0.8005–0.9013)0.720.8310.8720.6510.763 Combined0.866 (0.8182–0.9136)0.8480.7470.8420.7560.809Validation set (n=93) Clinic0.662 (0.5592–0.7639)0.8850.4630.6720.760.699 Intranodal0.739 (0.6356–0.8419)0.6150.8540.8420.6360.72 Habitat0.809 (0.7179–0.8994)0.6920.8540.8570.6860.763 Habitat H10.791 (0.6969–0.8848)0.6920.8050.8180.6730.742 Habitat H20.714 (0.6091–0.8182)0.5770.8540.8330.6140.699 Habitat H30.736 (0.6327–0.8396)0.750.6590.7360.6750.71 Habitat H40.787 (0.6912–0.8819)0.750.7560.7960.7050.753 Combined0.819 (0.7299–0.9080)0.6350.9270.9170.6670.763External test set (n=108) Clinic0.646 (0.6311–0.6618)0.9030.4450.6740.7820.701 Intranodal0.718 (0.7012–0.7345)0.6610.7930.8030.6480.719 Habitat0.791 (0.7769–0.8051)0.6960.7980.8140.6730.741 Habitat H10.688 (0.6708–0.7043)0.750.5940.7020.6520.682 Habitat H20.756 (0.7406–0.7719)0.5680.910.8890.6240.719 Habitat H30.665 (0.6480–0.6828)0.5350.7990.7720.5750.651 Habitat H40.731 (0.7149–0.7481)0.6310.8780.8680.6520.74 Combined0.806 (0.7918–0.8192)0.6290.8420.8350.6410.723AUC, area under the curve; CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value. Performance of the clinical, intranodal radiomics, habitat radiomics, and combined nomogram models in the training, internal validation, and external test cohorts AUC, area under the curve; CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value. Performance of combined nomogram In multivariable analysis, lymph node vascular pattern and cortical thickness were identified as independent risk factors associated with ALN metastasis. The 24 retained radiomics features were integrated into a single Habitat Rad-score. A combined nomogram was then constructed using the Habitat Rad-score, cortical thickness, and lymph node vascular pattern, resulting in three predictors, as shown in Figure 6. With 132 ALNM events in the training cohort, the EPV for the combined model was 44.00 (132/3). The combined nomogram exhibited the highest predictive performance in the training, internal validation, and external test cohorts, with AUC values of 0.866 (95% CI: 0.8182–0.9136), 0.819 (95% CI: 0.7299–0.9080), and 0.806 (95% CI: 0.7918–0.8192), respectively, as shown in Figure 7. The combined model significantly outperformed the clinical model (P<0.001, P=0.004, and P=0.02) and the intranodal radiomics model (P<0.001, P=0.02, and P=0.001) in the training, internal validation, and external test cohorts, respectively. However, its improvement over the habitat radiomics model was not statistically significant (P=0.40, P=0.54, and P=0.33, respectively; Table S1). Nomogram for the preoperative prediction of axillary lymph node metastasis. For vascular pattern, 0 indicates hilar vascularity and 1 indicates mixed/peripheral vascularity. For cortical thickness, 0 indicates <3 mm and 1 indicates ≥3 mm. ROC, calibration, and DCA of the different models in the training, internal validation, and external test cohorts. (A-C) The ROC curves, (D-F) the calibration curves, and (G-I) the DCA curves, corresponding to the training, internal validation, and external test cohorts, respectively. Intra denotes the intranodal radiomics model constructed from the entire intranodal ROI. AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; ROC, receiver operating characteristic; ROI, region of interest. Calibration curve analysis demonstrated good agreement between the predicted and observed probabilities of ALN metastasis. The Brier scores of the combined model were 0.146, 0.178, and 0.186 in the training, internal validation, and external test cohorts, respectively, indicating relatively stable probabilistic performance across cohorts. DCA further showed that the nomogram yielded a higher net benefit across most clinically relevant threshold probabilities. These results suggest that the combined nomogram has favorable predictive performance and may serve as a practical noninvasive tool for individualized preoperative evaluation of suspicious axillary lymph nodes. Discussion In this two-center study, we constructed a noninvasive habitat radiomics model based on two-dimensional ultrasound images of suspected axillary lymph nodes, and further combined clinical and ultrasound predictors to establish a combined nomogram. Unlike traditional intranodal radiomics that directly extracts features from the entire ROI, habitat analysis first performs unsupervised clustering based on pixel-level imaging features, partitioning each lymph node into four subregions with similar imaging phenotypes. Radiomics features from each subregion are then extracted and integrated, thereby more precisely characterizing the internal heterogeneity of lymph nodes. The results showed that habitat radiomics consistently outperformed the traditional intranodal radiomics in the training set, validation set and external test set. After further integrating lymph node vascular pattern and cortical thickness, the combined nomogram achieved the best performance, with AUC of 0.866, 0.819 and 0.806 in the training, validation and external test set, respectively. Contemporary axillary management increasingly tailors the extent of evaluation and treatment to individual patients. SOUND and INSEMA support the omission of SLNB in carefully selected clinically node-negative patients. These findings underscore the importance of accurate patient selection and individualized risk assessment. Although multicenter studies have confirmed that ultrasound-based radiomics and deep learning may improve preoperative prediction of ALNM (11,26), most have focused on the primary tumor or have extracted global features based on the ROI of the overall lymph nodes (27), without adequately revealing the spatial heterogeneity within the lymph node. Given the complex internal architecture of lymph nodes, we focused specifically on suspicious axillary lymph nodes, aiming to characterize intranodal microenvironmental heterogeneity through noninvasive imaging. This approach may complement conventional ultrasonography by providing quantitative information for individualized preoperative risk assessment in this selected population. This study confirmed the predictive value of intranodal radiomics for ALNM, yielding AUCs of 0.761, 0.739, and 0.718 in the training, validation, and external test sets, respectively. These results align well with the performance of a shear wave elastography-based radiomics model (28), which achieved C-indices of 0.845 and 0.817 in the training and validation sets. These findings suggest that radiomics based on feature extraction from the entire ROI is clinically feasible for assessing axillary lymph node status. However, further analysis revealed that intranodal radiomics model was relatively limited, with a specificity of only 0.566 in the training set. A possible reason is that radiomics features extracted from the entire ROI predominantly capture global texture patterns, thereby overlooking the spatial heterogeneity that may be critical for identifying subtle metastatic changes. To overcome this limitation, habitat analysis was further introduced in our study. The habitat radiomics models achieved AUCs of 0.859, 0.809, and 0.791 in the training, validation, and external test sets, respectively, with corresponding specificities of 0.867, 0.854, and 0.798. Compared with the intranodal radiomics model, the habitat radiomics model exhibited markedly higher specificity and greater stability across datasets. This improvement can be attributed to the way habitat analysis partitions the axillary lymph node into four subregions with distinct imaging phenotypes before feature extraction, thus preserving spatial heterogeneity that intranodal radiomics tends to overlook. A previous study showed that habitat analysis based on dynamic contrast-enhanced MRI achieved high accuracy in predicting ALNM in breast cancer (22). Unlike previous studies that mainly focused on overall habitat features, our study further divided the suspicious lymph node images into four independent habitat subregions and modeled for each subregion separately. The significant differences in predictive performance among the subregions suggest that distinct spatial regions within the same lymph node may carry distinct and complementary metastasis-related imaging information. Among the four independent subregions, H4 demonstrated relatively stable predictive ability across the training, validation, and external test sets (the AUCs were 0.851, 0.787 and 0.731 respectively), and its specificity was generally high (0.831, 0.756 and 0.878 respectively). Based on the visualization, H4 is predominantly distributed in a circular pattern around the periphery of the lymph nodes, suggesting that it may correspond to imaging changes in the cortical and marginal regions. Biologically, ALNM typically begins with tumor cells entering the marginal sinuses via afferent lymphatic vessels, followed by cortical remodeling and marginal infiltration. This process is indirectly supported by prior imaging evidence: Choi et al. identified cortical thickness >3 mm as a reliable sonographic predictor of lymph node metastasis (29); Loonis et al. and Kurt et al. further confirmed that asymmetric cortical thickening is an independent predictor of lymph node involvement (30,31). Therefore, we speculate that the features captured by H4 may be related to the cortical involvement and marginal structure remodeling in early metastatic progression. In contrast, Habitat H3 (blue area) showed excellent sensitivity (0.902 in the training cohort), but its external validation performance was comparatively limited. Considering its intermediate spatial location, H3 may correspond to a transition zone associated with tumor infiltration within the lymph node, or reflect active interactions between tumor cells and the paracortical host microenvironment. The paracortical area plays an important role in lymph node immune responses (17,32). The early microstructural changes associated with deep tumor infiltration may be more easily captured in this area, which may partly explain the higher sensitivity of H3. However, immune cell accumulation associated with reactive hyperplasia may produce imaging features similar to those of early metastatic involvement (33), thereby limiting the specificity of this subregion. In pathologically confirmed metastatic lymph nodes, Habitat H1 (red area) was mainly distributed in the central region of the lymph node. H1 may represent lymph nodes with advanced metastatic involvement, in which the medulla and fatty hilum are extensively replaced by tumor tissue. Habitat analysis provides a spatial framework for capturing regional heterogeneity within the lymph node. The integration of complementary spatial information may explain why the habitat radiomics model outperformed individual habitat radiomics and the intranodal radiomics model. Therefore, we believe that the 4-category habitat analysis approach adopted in this study is both interpretable and clinically relevant. However, the final correlation still needs to be confirmed by comparing corresponding histopathological images. Although the habitat radiomics model significantly improved predictive performance, lymph node metastasis involves not only subtle intranodal changes captured by radiomics, but also morphological and hemodynamic changes. Consistent with previous studies, features such as abnormal lymph node morphology and increased peripheral blood flow have been reported as predictors of N2–3 stage disease (34), particularly the presence of penetrant or anarchic vascular patterns (35). Accordingly, we incorporated clinical and ultrasound variables including age, orientation, quadrant, cortical thickness, lymph node vascular pattern, and molecular subtype. Multivariable analysis identified cortical thickness ≥3 mm (OR =3.436) and lymph node vascular pattern (OR =2.555) as independent predictors of ALN metastasis. After combining habitat features with these clinical and ultrasound factors, the nomogram achieved a specificity of 0.842 and an AUC of 0.806 in the external test set, supporting its potential clinical utility. This study has several limitations. First, although our study used LASSO regularization, ten-fold cross-validation, and independent external validation, the model may still be at risk of overfitting. Larger prospective multicenter studies are therefore needed. Second, the absence of spatially matched histopathologic validation limited the biological interpretation of the H1-H4 habitats. Future studies should combine spatially matched histopathologic and immunohistochemical analyses to clarify their biological basis. Third, the inclusion of only patients with sonographically suspicious axillary lymph nodes may limit the generalizability of our findings. Further multicenter studies in broader patient populations are needed to assess the model’s applicability in different clinical settings. Conclusions In conclusion, we developed and externally validated a visual nomogram for the preoperative prediction of ALNM in breast cancer by integrating a habitat-based ultrasound radiomics signature with clinical-ultrasound predictors. This two-center retrospective study demonstrated that ultrasound habitat radiomics of suspicious axillary lymph nodes better characterized intranodal heterogeneity and improved the assessment of metastatic status compared with intranodal radiomics. For future noninvasive evaluation of suspicious axillary lymph nodes, the nomogram models combining habitat radiomics with clinical and ultrasound factors may be prioritized. This approach provides a practical tool for individualized risk stratification and axillary management in breast cancer patients. Supplementary The article’s supplementary files as Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0363/prf