Integrating intratumoral and peritumoral ultrasound radiomics with clinicopathological features to predict axillary lymph node metastasis in early breast cancer

Introduction

Breast cancer (BC) remains the most frequently diagnosed malignancy and the leading cause of cancer-related mortality among women worldwide (1,2). In the management of early BC (EBC), axillary lymph node metastasis (ALNM) represents a pivotal determinant of pathological staging and is a key prognostic factor influencing prognosis (3). At present, sentinel lymph node biopsy (SLNB) is considered the standard approach for axillary staging in patients with clinically node-negative disease (cN0) (4-6). As the primary site of lymphatic drainage, the sentinel lymph node provides essential information regarding the overall axillary metastatic status. Despite its diagnostic value, SLNB is an invasive procedure and is associated with limitations that include postoperative complications such as shoulder dysfunction, nerve injury, sensory disturbances of the upper limb, and lymphedema, all of which may negatively affect quality of life. In addition, SLNB carries a false-negative rate, which may compromise the accuracy of clinical decision-making (7,8). With the continued improvement in early detection of BC, many patients now present with relatively low tumor burden, accompanied by a declining incidence of ALNM. Consequently, the clinical benefit of SLNB is diminished in some patients, reinforcing the need to develop accurate, non-invasive, preoperative methods for assessing ALNM.

Ultrasound is a widely used imaging modality for the preoperative evaluation of breast lesions and remains the primary non-invasive technique for assessing axillary lymph node status (9,10). However, in most patients with EBC with cN0, axillary ultrasound fails to detect suspicious lymph nodes (11,12). Although previous studies have reported associations between imaging features of primary breast tumors and ALNM, conventional image interpretation relies heavily on subjective assessment by radiologists, resulting in limited and variable diagnostic performance (13). Radiomics offers a promising solution to this challenge. As a quantitative imaging technique, radiomics extracts high-dimensional features from conventional imaging data using computational algorithms, thereby uncovering information beyond visual perception. By revealing associations between imaging phenotypes and biological behavior, radiomics enhances diagnostic precision, prognostic assessment, and individualized treatment planning (14,15). In recent years, accumulating evidence has demonstrated that intratumoral radiomics features derived from breast imaging can achieve favorable performance in predicting ALNM (16-18). Nevertheless, most existing radiomics studies have focused exclusively on intratumoral regions, largely overlooking the potential contribution of the peritumoral region, which contains critical microenvironmental information such as vascular networks, stromal alterations, and inflammatory changes that may influence metastatic potential. Accordingly, integrating intratumoral and peritumoral radiomics features may further enhance the accuracy of ALNM prediction. Concurrently, advances in artificial intelligence (AI), particularly the application of machine learning (ML) algorithms to radiomics data, have markedly enhanced predictive performance and clinical applicability (19). Given these developments, this study aimed to develop and validate an ML-based predictive model incorporating intratumoral and peritumoral ultrasound radiomics features with clinicopathological variables, with Shapley additive explanations (SHAP) integrated to enhance interpretability and elucidate individual feature contributions, thereby providing a more accurate, transparent, and non-invasive preoperative tool for ALNM assessment in patients with EBC. We present this article in accordance with the TRIPOD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0015/rc).

Methods

Patient selection

Patients who underwent surgical treatment and received a pathological diagnosis of BC at Jiaxing Maternity and Child Health Care Hospital between September 1, 2021, and June 30, 2025, were retrospectively screened for eligibility. Preoperative breast ultrasound images and complete clinicopathological data were collected for all eligible cases. After applying the inclusion and exclusion criteria, a total of 327 patients with EBC, aged 28–81 years, were included in the final analysis. The study population was randomly divided into a training cohort (n=228) and a validation cohort (n=99) at a 7:3 ratio. The training cohort was used for radiomics features selection and predictive model development, whereas the validation cohort was used for evaluating model generalizability. The overall study workflow is illustrated in Figure 1. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Jiaxing Maternity and Child Health Care Hospital (approval No. 2025-Y-106). As this study employed a retrospective analysis design, the requirement for individual informed consent was waived.

Figure 1 Flow diagram of the study.

Inclusion and exclusion criteria

Patients were included if they met the following criteria: (I) diagnosis of EBC with cN0 and without distant metastasis, with a maximal diameter ≤5 cm; (II) presence of a solitary breast lesion confirmed by imaging; (III) absence of prior neoadjuvant therapy or breast surgery; (IV) performance of intraoperative SLNB using methylene blue as a single tracer to ensure consistent localization; and (V) availability of high-quality ultrasound images suitable for radiomics analysis along with complete clinicopathological data.

Patients were excluded if (I) ultrasound images exhibited significant artifacts or insufficient resolution that precluded accurate region-of-interest (ROI) delineation; (II) concomitant other primary malignancies were present; (III) lesions were non-mass-like with indistinct boundaries precluding accurate tumor delineation; or (IV) clinicopathological data were incomplete or key variables were missing.

Clinicopathological features

Clinicopathological variables were systematically extracted from the institutional electronic medical record system. Collected data included patient baseline characteristics such as age, body mass index, and menstrual status, as well as tumor-related variables including maximal diameter, tumor location, lymphovascular invasion status, and histological type. Histological grading was performed according to the Nottingham grading system and categorized as Grades I–III. Patients were classified into ALNM-positive and ALNM-negative groups based on SLNB findings. Estrogen receptor (ER) and progesterone receptor (PR) status were defined as positive when ≥1% of tumor cell nuclei demonstrated immunohistochemical staining, and cases meeting this threshold were classified as hormone receptor (HR) positive (20). Human epidermal growth factor receptor 2 (HER2) status was determined by immunohistochemistry (scores 0–3+), with tumors with a HER2 score of 2+ undergoing fluorescence in situ hybridization (21). The Ki67 proliferation index was categorized as low (<30%) or high (≥30%). Based on immunohistochemical profiles, BCs were categorized into three molecular subtypes: luminal, HER2-positive, and triple-negative BC subtypes.

Ultrasound image acquisition and delineation

All ultrasound examinations were performed by physicians with >5 years of experience in breast ultrasound diagnosis to ensure reliability. Patients were positioned in the supine position with both arms elevated to optimize lesion visualization. Comprehensive multiplanar scanning was performed, and images capturing the maximum long- and short axes of the lesion were acquired and stored in Digital Imaging and Communications in Medicine (DICOM) format.

To address potential variability across ultrasound devices, image intensity normalization was performed before ROI delineation. ROI were delineated using ITK-SNAP with meticulous attention to morphological features such as spiculated margins, angular contours, and lobulated architecture. In cases where lesion boundaries were indistinct, two senior radiologists jointly reviewed and reached a consensus to ensure contour accuracy.

Based on prior studies by Cui et al. (22) and Ding et al. (23), we delineated three types of ROIs in this study: intratumoral ROI, 4-mm peritumoral ROI, and combined intratumoral--peritumoral ROI (Figure 2). The 4-mm margin was selected because it captures the biologically relevant peritumoral microenvironment at the invasive tumor front while remaining within the breast parenchyma for most lesions, minimizing inclusion of distant normal tissue that could dilute tumor-specific radiomic signals.

Figure 2 Using ITK-SNAP software to delineate ROI. (A) Original two-dimensional grayscale ultrasound image; (B) intratumoral ROI; (C) peritumoral ROI; (D) intratumoral and peritumoral ROI. ROI, region of interest.

Radiomics features extraction and selection

Radiomics features were extracted systematically from standardized ultrasound images to ensure detailed characterization of both tumor morphology and textural heterogeneity, using the PyRadiomics library in Python. The full feature set encompassed seven categories: Shape, First-Order Statistics, Gray-Level Co-occurrence Matrix, Gray-Level Run-Length Matrix, Gray-Level Size Zone Matrix, Gray-Level Dependence Matrix, and Neighboring Gray-Tone Difference Matrix.

To develop a robust and clinically interpretable predictive model, a multi-step feature optimization pipeline was implemented. Initially, outlier detection was performed to remove extreme values and ensure stability during subsequent Z-score standardization. All extracted features then underwent Z-score normalization to achieve uniform scaling. Spearman correlation analysis was subsequently conducted to identify highly correlated features, and features with correlation coefficients >0.9 were eliminated to reduce multicollinearity. Recursive feature elimination was then applied to iteratively remove features, further refining the dataset. Finally, least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation was applied to identify the most informative features. The optimal regularization parameter (lambda) was selected by minimizing the mean squared error, and only features with non-zero coefficients at the optimal lambda were retained for model construction.

Predictive model construction and performance evaluation

Radiomics model development

Following feature selection, radiomics scores (rad-scores) were calculated for each patient by applying weighted linear coefficients. Three established ML algorithms—logistic regression (LR), support vector machine (SVM), and random forest (RF)—were employed to construct radiomics models based on intratumoral, peritumoral, and combined intratumoral-peritumoral features. This approach enabled a methodical comparison of predictive performance across different anatomical regions.

Model performance was evaluated using receiver operating characteristic (ROC) curve analysis, with the area under the curve (AUC) serving as the primary metric of discriminative capability. Additional performance metrics, including accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), were also calculated, allowing detailed characterization of model behavior across distinct clinical scenarios.

Clinical-radiomics combined model

Clinicopathological variables that demonstrated statistical significance in univariate analysis (P<0.05) were entered into multivariate LR to identify independent predictors of ALNM. These independent clinical predictors were subsequently integrated with rad-scores derived from the optimal radiomics model to construct a combined clinical-radiomics prediction model. Model performance was evaluated in both the training and validation cohorts using ROC curve analysis, and DeLong tests were applied to compare differences in AUC between models (24). Calibration performance was evaluated using calibration curves and the Hosmer-Lemeshow test, where P>0.05 indicated acceptable calibration. Decision curve analysis (DCA) quantified net clinical benefit across varying threshold probabilities, assessing whether the model would provide meaningful clinical utility beyond treat-all or treat-none strategies (25).

Interpretability analysis

To enhance model transparency and clinical interpretability, the optimal predictive model was further analyzed using the SHAP framework (26,27). This method, grounded in cooperative game theory, quantifies how individual feature contributions influence the predicted probability of ALNM and provides both global and patient-level interpretability.

Statistical analysis

All statistical analyses were conducted using SPSS version 29.0 (SPSS Inc., Chicago, IL, USA) and Python version 3.9 (http://www.python.org). Categorical variables are reported as frequencies and percentages, and comparisons between groups were performed using the chi-square test. All statistical tests were two-tailed, and a P value <0.05 was considered to indicate statistical significance.

Results

Clinicopathological features selection

A total of 327 patients with EBC were included in this study, of whom 152 were diagnosed with ALNM and 175 were ALNM-negative, yielding an overall metastasis incidence of 46.48%. The incidence of ALNM was comparable between the training and validation cohorts (46.49% vs. 46.46%, respectively), and no significant differences were observed in baseline clinicopathological features between the two cohorts (P>0.05; Table 1). Univariate analysis identified multiple potential predictors associated with ALNM (Table 2), which were subsequently entered into multivariate LR analysis. This analysis identified maximal diameter and lymphovascular invasion as independent predictors of ALNM (Table 3).

Radiomics features extraction and selection

A total of 469 intratumoral, 469 peritumoral, and 938 combined intratumoral-peritumoral radiomics features were extracted. Stepwise dimensionality reduction followed by LASSO regression (Figure 3) resulted in the selection of 11 intratumoral, 10 peritumoral, and 22 combined-region features (Figure 4). Pearson correlation heatmaps confirmed that selected features demonstrated minimal multicollinearity (Figure 5), enhancing the stability and generalizability of downstream predictive models.

Figure 3 LASSO regression for ultrasound radiomics features screening. (A,C,E) Characteristic coefficient changing curve with Lambda. (B,D,F) Lambda selection graph. LASSO, least absolute shrinkage and selection operator.

Figure 4 The optimal radiomics features selected by LASSO regression. LASSO, least absolute shrinkage and selection operator.

Figure 5 Heatmap of correlations among the optimal radiomics features.

Radiomics models performance comparison

To evaluate the predictive value of radiomics features derived from different anatomical regions, rad-scores were calculated based on the selected features and assessed using LR, SVM, and RF algorithms (Table 4).

For the intratumoral radiomics model, the RF algorithm demonstrated the best performance in the training cohort, achieving an AUC of 0.903 [95% confidence interval (CI): 0.864–0.940], with an accuracy of 0.829, sensitivity of 0.736, and specificity of 0.910. However, in the validation cohort, the LR model exhibited superior generalization performance, with an AUC of 0.813 (95% CI: 0.733–0.892), suggesting better generalization. The RF algorithm exhibited performance deterioration in the validation cohort (AUC =0.760), indicating possible overfitting (Figure 6A,6B).

Figure 6 Comparison of ROC curves of different models in the training cohort and validation cohort. AUC, area under the curve; LR, logistic regression; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine.

For the peritumoral radiomics model, the RF algorithm again achieved the highest discriminative ability in the training cohort, with an AUC of 0.845 (95% CI: 0.795–0.893). The LR model demonstrated the best performance in the validation cohort (AUC =0.787; 95% CI: 0.698–0.874), with consistent accuracy and sensitivity metrics. The SVM algorithm showed weaker predictive performance (AUC =0.747; Figure 6C,6D).

The combined intratumoral-peritumoral radiomics model demonstrated the best overall performance among all modeling strategies. In the training cohort, the LR and RF models achieved AUCs of 0.890 and 0.900, respectively. In the validation cohort, the LR algorithm yielded the highest AUC of 0.868 (95% CI: 0.795–0.933), improving predictive accuracy by 6.8% compared with the intratumoral-only model and 10.3% compared with the peritumoral-only model (Figure 6E,6F). These findings support the added predictive value of multi-region radiomics feature integration.

Construction and validation of clinical-radiomics combined model

Based on feature selection, the two independent clinical predictors—maximal diameter and lymphovascular invasion—were integrated with the combined-region rad-scores to construct three ML-based combined models (Table 4). The LR combined model demonstrated the strongest performance in the training cohort (AUC =0.920; 95% CI: 0.882–0.951), with balanced accuracy, sensitivity, specificity, PPV, and NPV. The RF and SVM models exhibited slightly lower performance, and the SVM model showed suboptimal sensitivity. Validation cohort analysis further confirmed the LR combined model as the most reliable, with an AUC of 0.909 (95% CI: 0.851–0.960) and strong performance across all metrics (Figure 7A,7B). Based on the above analysis, LR algorithm was adopted to systematically compare five prediction models (Table 5). The AUC of each model in the training cohort showed an increasing trend, with the combined model achieving optimal performance. Validation cohort evaluation further confirmed the excellent generalization ability of the combined model. In contrast, the clinical-only model exhibited a sensitivity of only 0.478, raising concerns regarding potential missed diagnoses (Figure 8A,8B). DeLong test results demonstrated that the predictive performance of the combined model was significantly superior to that of the other models (P<0.05) (Table 6). Calibration analyses demonstrated excellent agreement between predicted and observed outcomes, supported by Hosmer-Lemeshow test results and calibration curves (Figure 8C). DCA demonstrated a clear net clinical benefit across a wide threshold range, confirming strong clinical applicability (Figure 8D,8E).

Figure 7 ROC curve performance comparison of the clinical-radiomics combined model between the training and validation cohorts. AUC, area under the curve; LR, logistic regression; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine.

Figure 8 Development and validation of the clinical-radiomics combined model. (A,B) ROC curves of different models based on logistic regression algorithm in the training cohort (A) and validation cohort (B); (C) Calibration curves of the clinical-radiomics combined model in the training and validation cohort; (D,E) DCA of different models in the training cohort (D) and validation cohort (E). Model 1: clinical model; Model 2: intratumoral radiomics model; Model 3: peritumoral radiomics model; Model 4: combined intratumoral-peritumoral radiomics model; Model 5: clinical-radiomics combined model. AUC, area under the curve; DCA, decision curve analysis; ROC, receiver operating characteristic.

SHAP explainability analysis

To elucidate the decision-making process of the LR-based clinical-radiomics combined model, SHAP analysis was performed. Global SHAP interpretation showed that the rad-scores derived from the combined ROI contributed most substantially to model predictions, exceeding the influence of traditional clinicopathological variables (Figure 9A). This result reinforces the value of radiomics features in ALNM prediction. Individual-level SHAP force plots for representative cases with positive and negative ALNM (Figure 9B,9C) illustrated both the direction and magnitude of each feature’s contribution to the final prediction, with positive SHAP values pushing predictions toward ALNM positivity, while negative values contributed to negative classifications. These individualized prediction pathways enhance the interpretability and clinical transparency of the model.

Figure 9 Model interpretation by SHAP. (A) The overall contribution of each feature to the model prediction classification (class 0: non-response, class 1: response), in order of importance. (B) SHAP force plot of patients with ALNM. (C) SHAP force plot of patients with ALNM-negative. ALNM, axillary lymph node metastasis; SHAP, Shapley additive explanations.

Discussion

With the evolution of precision oncology, SLNB has progressively replaced conventional axillary lymph node dissection as the gold standard for axillary staging, reducing surgical morbidity (29). However, epidemiological data indicate that over 70% of patients with EBC with cN0 do not harbor ALNM, raising increasing concern regarding the routine application of SLNB universally in this group (22,30). In this context, the development of accurate preoperative prediction models to identify low-risk patients could meaningfully reduce unnecessary invasive procedures and optimize individualized treatment pathways.

Ultrasound is a first-line imaging modality for axillary assessment owing to its non-invasive nature, cost-effectiveness, convenience, and high reproducibility (31,32). Nevertheless, early-stage ALNM does not always produce conspicuous nodal morphological changes, limiting the sensitivity for conventional sonographic staging (33,34). This limitation has motivated increasing research into whether imaging features of the primary breast lesions correlate with ALNM and whether computational approaches can extract predictive signals beyond human interpretation (13,35). With the rapid advancement of AI, ALNM prediction studies combining clinicopathological features with radiomics features have emerged, demonstrating remarkable predictive performance. Yu et al. (36) developed an ultrasound-based radiomics model for EBC that achieved AUCs of 0.780 and 0.710 in the training and validation cohorts, respectively, indicating moderate predictive capability. Similarly, Qiu et al. (37) reported AUCs of 0.778 and 0.725 using a model based on 21 ultrasound features. Zhou et al. (38) further demonstrated that a radiomics-based model significantly outperformed experienced physicians in ALNM prediction, achieving an AUC of 0.850 compared with 0.590 for radiologist assessment (P<0.01). Importantly, most of these studies have focused exclusively on intratumoral features and have not fully interrogated the peritumoral microenvironment, which is biologically relevant under the “seed and soil” paradigm of metastasis. Recent evidence supports the additive value of peritumoral imaging. For example, Sun et al. (39) showed that integrating a 5-mm peritumoral radiomics signature substantially improved AUC compared with intratumoral features alone, demonstrating the peritumoral region’s contribution to metastasis prediction. Building on this rationale, the present study systematically constructed and validated multiple ML models based on intratumoral, peritumoral 4-mm, and combined intratumoral-peritumoral radiomics features, integrated with clinicopathological predictors to develop a clinical-radiomics fusion model for ALNM prediction.

Multivariate LR analysis identified maximal diameter and lymphovascular invasion as independent predictors of ALNM, aligning with established prognostic literature (40-44). Lymphovascular invasion emerged as the strongest clinical predictor, consistent with its biological role in facilitating tumor cell entry into lymphovascular channels. Maximal diameter was correlated with intrinsic tumor aggressiveness, angiogenesis, and intratumoral heterogeneity—mechanistic factors that promote metastatic dissemination. Although the clinicopathological model achieved acceptable overall discrimination, its low sensitivity highlighted the inherent limitations of relying on clinicopathological variables alone for reliable preoperative ALNM screening. In contrast, radiomics-based modeling substantially improved predictive capability. The intratumoral LR model supported the predictive value of intratumoral texture and shape descriptors. Although the peritumoral LR model showed comparatively lower discriminatory performance, its biological significance remains important, as peritumoral alterations likely reflect tumor-host interactions such as inflammation, vascular remodeling, and extracellular matrix remodeling, which are upstream processes in metastatic progression (45-48). When combined with the two independent clinical predictors (maximal diameter and lymphovascular invasion), the LR clinical-radiomics model achieved excellent performance, and improved sensitivity by nearly 70% relative to the clinical model alone, a clinically meaningful gain that could reduce missed metastatic cases and may facilitate more informed surgical decision-making and axillary management. Algorithm selection influenced generalizability: although RF achieved higher training AUCs, LR consistently exhibited superior generalization in the validation cohort, suggesting lower overfitting risk for LR in our dataset. This observation likely reflects the favorable bias-variance characteristics of regularized LR when applied to radiomics feature sets with reduced dimensionality and moderate sample sizes. We therefore adopted LR for final model comparison and potential clinical deployment. SHAP explainability analysis clarified the model’s decision logic: the combined rad-scores derived from intratumoral and peritumoral features was the most influential predictor, while clinicopathological variables contributed additively. This finding both corroborates the predictive utility of radiomics and provides clinically interpretable evidence that the radiomics signature captures relevant biological signal. The availability of SHAP visualizations for individual patients further enhances transparency and facilitates clinician acceptance by showing how specific features drive predictions. Compared with prior ultrasound-based radiomics studies that achieved validation AUCs of 0.710–0.850 using intratumoral features alone, our model integrates both intratumoral and 4-mm peritumoral signatures with independent clinical predictors to attain a substantially higher validation AUC of 0.909, underscoring the additive value of peritumoral microenvironmental information that prior approaches have not fully exploited (36-38). Furthermore, unlike these existing studies that offered limited model transparency, our SHAP-based interpretability framework quantifies individual feature contributions.

This study has several limitations. First, this was a single-center retrospective study with a modest sample size and a relatively small validation cohort, which may introduce selection bias and constrain external validity. Multicenter external validation is required to confirm model robustness across diverse ultrasound systems, operator protocols, and patient populations. Second, the 4-mm peritumoral margin was chosen based on prior reports; however, the optimal peritumoral expansion may vary across lesions and modalities. Future work should investigate a range of peritumoral distances and potentially adopt adaptive peritumoral definitions informed by tumor size, histology, or anatomical context. Third, ROI delineation inherently depends on operator judgment and intensity normalization; automated or semi-automated segmentation with interobserver reproducibility analysis would strengthen methodological rigor. Fourth, while traditional ML algorithms achieved satisfactory performance, future research may benefit from exploring deep learning approaches for their potential to learn hierarchical features directly from images, particularly as larger multicenter datasets become available. Additionally, lymphovascular invasion was retrospectively determined from postoperative pathology, which may be subject to sampling and interobserver variability; prospective validation through standardized intraoperative specimen collection is warranted to confirm its reproducibility as a robust predictive factor.

Conclusions

We successfully developed and validated an LR-based clinical-radiomics model that integrates intratumoral and 4-mm peritumoral ultrasound radiomics signatures with clinical predictors (maximal diameter and lymphovascular invasion) as a non-invasive tool for preoperative assessment of ALNM in patients with EBC. The combined model demonstrated strong discrimination, robust calibration, and favorable clinical net benefit in DCA, outperforming single-modality models and a clinicopathological model alone. To translate these promising findings into clinical practice, prospective implementation studies are warranted to determine whether the model can safely reduce SLNB rates without compromising oncologic outcomes.

Acknowledgments

We appreciate the linguistic assistance provided by TopEdit (www.topeditsci.com) during the preparation of this manuscript.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0015/rc

Data Sharing Statement: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0015/dss

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0015/prf

Funding: This work was supported by the Jiaxing Health Science and Technology Plan Project (No. JWKZ-25004).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0015/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. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Jiaxing Maternity and Child Health Care Hospital (approval No. 2025-Y-106). As this study employed a retrospective analysis design, the requirement for individual informed consent was waived.

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. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Siegel RL, Miller KD, Wagle NS, et al. Cancer statistics, 2023. CA Cancer J Clin 2023;73:17-48. [Crossref] [PubMed]
3. Andring LM, Diao K, Sun S, et al. Locoregional Management and Prognostic Factors in Breast Cancer With Ipsilateral Internal Mammary and Axillary Lymph Node Involvement. Int J Radiat Oncol Biol Phys 2022;113:552-60. [Crossref] [PubMed]
4. Heerdt AS. Lymphatic Mapping and Sentinel Lymph Node Biopsy for Breast Cancer. JAMA Oncol 2018;4:431. [Crossref] [PubMed]
5. Lyman GH, Giuliano AE, Somerfield MR, et al. American Society of Clinical Oncology guideline recommendations for sentinel lymph node biopsy in early-stage breast cancer. J Clin Oncol 2005;23:7703-20. [Crossref] [PubMed]
6. Valente SA, Levine GM, Silverstein MJ, et al. Accuracy of predicting axillary lymph node positivity by physical examination, mammography, ultrasonography, and magnetic resonance imaging. Ann Surg Oncol 2012;19:1825-30. [Crossref] [PubMed]
7. Kootstra JJ, Hoekstra-Weebers JE, Rietman JS, et al. A longitudinal comparison of arm morbidity in stage I-II breast cancer patients treated with sentinel lymph node biopsy, sentinel lymph node biopsy followed by completion lymph node dissection, or axillary lymph node dissection. Ann Surg Oncol 2010;17:2384-94. [Crossref] [PubMed]
8. Langer I, Guller U, Berclaz G, et al. Morbidity of sentinel lymph node biopsy (SLN) alone versus SLN and completion axillary lymph node dissection after breast cancer surgery: a prospective Swiss multicenter study on 659 patients. Ann Surg 2007;245:452-61. [Crossref] [PubMed]
9. Choi HY, Park M, Seo M, et al. Preoperative Axillary Lymph Node Evaluation in Breast Cancer: Current Issues and Literature Review. Ultrasound Q 2017;33:6-14. [Crossref] [PubMed]
10. Lee MC, Joh JE, Chau A. Axillary staging prior to neoadjuvant chemotherapy: the roles of sentinel lymph node biopsy and axillary ultrasonography. Cancer Control 2012;19:277-85. [Crossref] [PubMed]
11. Riedel F, Schaefgen B, Sinn HP, et al. Diagnostic accuracy of axillary staging by ultrasound in early breast cancer patients. Eur J Radiol 2021;135:109468. [Crossref] [PubMed]
12. Rukanskienė D, Veikutis V, Jonaitienė E, et al. Preoperative Axillary Ultrasound versus Sentinel Lymph Node Biopsy in Patients with Early Breast Cancer. Medicina (Kaunas) 2020;56:127. [Crossref] [PubMed]
13. Guo Q, Dong Z, Zhang L, et al. Ultrasound Features of Breast Cancer for Predicting Axillary Lymph Node Metastasis. J Ultrasound Med 2018;37:1354-3. [Crossref] [PubMed]
14. Conti A, Duggento A, Indovina I, et al. Radiomics in breast cancer classification and prediction. Semin Cancer Biol 2021;72:238-50. [Crossref] [PubMed]
15. Lambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012;48:441-6. [Crossref] [PubMed]
16. Fujioka T, Mori M, Kubota K, et al. The Utility of Deep Learning in Breast Ultrasonic Imaging: A Review. Diagnostics (Basel) 2020;10:1055. [Crossref] [PubMed]
17. Gao Y, Luo Y, Zhao C, et al. Nomogram based on radiomics analysis of primary breast cancer ultrasound images: prediction of axillary lymph node tumor burden in patients. Eur Radiol 2021;31:928-37. [Crossref] [PubMed]
18. Gu J, Jiang T. Ultrasound radiomics in personalized breast management: Current status and future prospects. Front Oncol 2022;12:963612. [Crossref] [PubMed]
19. Greener JG, Kandathil SM, Moffat L, et al. A guide to machine learning for biologists. Nat Rev Mol Cell Biol 2022;23:40-55. [Crossref] [PubMed]
20. Allison KH, Hammond MEH, Dowsett M, et al. Estrogen and Progesterone Receptor Testing in Breast Cancer: ASCO/CAP Guideline Update. J Clin Oncol 2020;38:1346-66. [Crossref] [PubMed]
21. Wolff AC, Hammond ME, Hicks DG, et al. Recommendations for human epidermal growth factor receptor 2 testing in breast cancer: American Society of Clinical Oncology/College of American Pathologists clinical practice guideline update. Arch Pathol Lab Med 2014;138:241-56. [Crossref] [PubMed]
22. Cui X, Wang N, Zhao Y, et al. Preoperative Prediction of Axillary Lymph Node Metastasis in Breast Cancer using Radiomics Features of DCE-MRI. Sci Rep 2019;9:2240. [Crossref] [PubMed]
23. Ding J, Chen S, Serrano Sosa M, et al. Optimizing the Peritumoral Region Size in Radiomics Analysis for Sentinel Lymph Node Status Prediction in Breast Cancer. Acad Radiol 2022;29:S223-8. [Crossref] [PubMed]
24. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 1988;44:837-45.
25. Kerr KF, Brown MD, Zhu K, et al. Assessing the Clinical Impact of Risk Prediction Models With Decision Curves: Guidance for Correct Interpretation and Appropriate Use. J Clin Oncol 2016;34:2534-40. [Crossref] [PubMed]
26. Chowdhury SU, Sayeed S, Rashid I, et al. Shapley-Additive-Explanations-Based Factor Analysis for Dengue Severity Prediction using Machine Learning. J Imaging 2022;8:229. [Crossref] [PubMed]
27. Lundberg SM, Lee SI. A unified approach to interpreting model predictions. Adv Neural Inf Process Syst 2017;30:4766-77.
28. Vohra LM, Gulzar R, Saleem O. INTRA OPERATIVE FROZEN EXAMINATION OF SENTINEL LYMPH NODE IN BREAST CANCER. J Ayub Med Coll Abbottabad 2015;27:40-4.
29. Malter W, Hellmich M, Badian M, et al. Factors Predictive of Sentinel Lymph Node Involvement in Primary Breast Cancer. Anticancer Res 2018;38:3657-62. [Crossref] [PubMed]
30. Reimer T, Engel J, Schmidt M, et al. Is Axillary Sentinel Lymph Node Biopsy Required in Patients Who Undergo Primary Breast Surgery? Breast Care (Basel) 2018;13:324-30. [Crossref] [PubMed]
31. Marino MA, Avendano D, Zapata P, et al. Lymph Node Imaging in Patients with Primary Breast Cancer: Concurrent Diagnostic Tools. Oncologist 2020;25:e231-42. [Crossref] [PubMed]
32. Qiu SQ, Zeng HC, Zhang F, et al. A nomogram to predict the probability of axillary lymph node metastasis in early breast cancer patients with positive axillary ultrasound. Sci Rep 2016;6:21196. [Crossref] [PubMed]
33. Cools-Lartigue J, Meterissian S. Accuracy of axillary ultrasound in the diagnosis of nodal metastasis in invasive breast cancer: a review. World J Surg 2012;36:46-54. [Crossref] [PubMed]
34. Ibrahim-Zada I, Grant CS, Glazebrook KN, et al. Preoperative axillary ultrasound in breast cancer: safely avoiding frozen section of sentinel lymph nodes in breast-conserving surgery. J Am Coll Surg 2013;217:7-15; discussion 15-6. [Crossref] [PubMed]
35. Bae MS, Shin SU, Song SE, et al. Association between US features of primary tumor and axillary lymph node metastasis in patients with clinical T1-T2N0 breast cancer. Acta Radiol 2018;59:402-8. [Crossref] [PubMed]
36. Yu FH, Wang JX, Ye XH, et al. Ultrasound-based radiomics nomogram: A potential biomarker to predict axillary lymph node metastasis in early-stage invasive breast cancer. Eur J Radiol 2019;119:108658. [Crossref] [PubMed]
37. Qiu X, Jiang Y, Zhao Q, et al. Could Ultrasound-Based Radiomics Noninvasively Predict Axillary Lymph Node Metastasis in Breast Cancer? J Ultrasound Med 2020;39:1897-905. [Crossref] [PubMed]
38. Zhou WJ, Zhang YD, Kong WT, et al. Preoperative prediction of axillary lymph node metastasis in patients with breast cancer based on radiomics of gray-scale ultrasonography. Gland Surg 2021;10:1989-2001. [Crossref] [PubMed]
39. Sun Q, Lin X, Zhao Y, et al. Deep Learning vs. Radiomics for Predicting Axillary Lymph Node Metastasis of Breast Cancer Using Ultrasound Images: Don't Forget the Peritumoral Region. Front Oncol 2020;10:53.
40. Chen H, Meng X, Hao X, et al. Correlation Analysis of Pathological Features and Axillary Lymph Node Metastasis in Patients with Invasive Breast Cancer. J Immunol Res 2022;2022:7150304. [Crossref] [PubMed]
41. Ding J, Jiang L, Wu W. Predictive Value of Clinicopathological Characteristics for Sentinel Lymph Node Metastasis in Early Breast Cancer. Med Sci Monit 2017;23:4102-8. [Crossref] [PubMed]
42. Lale A, Yur M, Özgül H, et al. Predictors of non-sentinel lymph node metastasis in clinical early stage (cT1-2N0) breast cancer patients with 1-2 metastatic sentinel lymph nodes. Asian J Surg 2020;43:538-49. [Crossref] [PubMed]
43. Li N, Song C, Huang X, et al. Optimized Radiomics Nomogram Based on Automated Breast Ultrasound System: A Potential Tool for Preoperative Prediction of Metastatic Lymph Node Burden in Breast Cancer. Breast Cancer (Dove Med Press) 2023;15:121-32. [Crossref] [PubMed]
44. Orsaria P, Caredda E, Genova F, et al. Additional Nodal Disease Prediction in Breast Cancer with Sentinel Lymph Node Metastasis Based on Clinicopathological Features. Anticancer Res 2018;38:2109-17. [Crossref] [PubMed]
45. Conklin MW, Keely PJ. Why the stroma matters in breast cancer: insights into breast cancer patient outcomes through the examination of stromal biomarkers. Cell Adh Migr 2012;6:249-60. [Crossref] [PubMed]
46. Lee AK, DeLellis RA, Silverman ML, et al. Prognostic significance of peritumoral lymphatic and blood vessel invasion in node-negative carcinoma of the breast. J Clin Oncol 1990;8:1457-65. [Crossref] [PubMed]
47. Liu Z, Chen J, Ren Y, et al. Multi-stage mechanisms of tumor metastasis and therapeutic strategies. Signal Transduct Target Ther 2024;9:270. [Crossref] [PubMed]
48. Ocaña A, Diez-Gónzález L, Adrover E, et al. Tumor-infiltrating lymphocytes in breast cancer: ready for prime time? J Clin Oncol 2015;33:1298-9. [Crossref] [PubMed]