Ultrasound radiomics for preoperative prediction of high axillary nodal burden in node-positive breast cancer: comparison of intratumoral and peritumoral features with external validation

Introduction

Breast cancer remains the most frequently diagnosed malignancy among women worldwide and is a leading cause of cancer-related mortality (1,2). Axillary lymph node (ALN) involvement is a key determinant of tumor staging, prognosis, and treatment planning (3). Beyond simple nodal positivity, accurate preoperative assessment of axillary nodal burden has become increasingly important because the extent of nodal involvement directly influences axillary surgical strategies and adjuvant treatment planning. However, conventional preoperative methods, including axillary ultrasound and needle biopsy, still have limited ability to accurately characterize the overall nodal burden (3). Therefore, a reliable and noninvasive approach for preoperative prediction of ALN burden remains clinically desirable.

Radiomics has emerged as a quantitative imaging technique that converts medical images into high-dimensional quantitative features, thereby providing information beyond conventional visual assessment (4,5). In breast cancer imaging, ultrasound-based radiomics has demonstrated potential in several applications, including lesion characterization, molecular subtype prediction, treatment response evaluation, and assessment of ALN status (6,7). Numerous studies have investigated the role of imaging and radiomics in predicting ALN metastasis in breast cancer. However, comparatively fewer studies have specifically focused on the preoperative prediction of axillary nodal burden, which may be more clinically relevant for guiding surgical decision-making. For example, Wang et al. developed a machine learning model integrating clinical variables with intratumoral and peritumoral ultrasound radiomics features to predict ALN metastasis in early-stage breast cancer (8). Subsequently, Wang et al. further extended this strategy to predict axillary nodal burden using ultrasound radiomics signatures (9). These findings suggest that ultrasound-based radiomics may serve as a promising tool for noninvasive preoperative nodal risk stratification.

Despite these encouraging results, several important issues remain to be clarified. First, most previous ultrasound radiomics studies have focused on predicting the presence or absence of ALN metastasis rather than the extent of nodal involvement, although nodal burden may be more clinically relevant for guiding surgical management (6,7,10-12). Second, many published models were developed using single-center datasets with internal random splitting, which may overestimate model performance and limit generalizability compared with external validation studies (13,14). Third, increasing attention has been given to the peritumoral region because it may reflect tumor-stroma interactions and microenvironmental heterogeneity. However, the incremental value of peritumoral ultrasound radiomics compared with intratumoral radiomics remains uncertain. Although several studies have reported promising results using peritumoral features, their robustness across different patient cohorts and their contribution to predicting ALN burden have not been fully established (8,15,16).

Therefore, the present study aimed to develop and externally validate ultrasound radiomics models for predicting high axillary nodal burden among node-positive breast cancer patients. Specifically, we compared intratumoral and 3-mm peritumoral radiomics signatures and evaluated clinical, radiomics, and combined models under a leakage-free external validation framework. By assessing discrimination, calibration, and decision curve analysis (DCA), this study sought to clarify the practical value and limitations of intratumoral and peritumoral ultrasound radiomics for preoperative nodal burden risk stratification. We present this article in accordance with the TRIPOD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0226/rc).

Methods

Study population

This retrospective dual-center study included patients with pathologically confirmed primary breast cancer who underwent preoperative ultrasound examination between January 2016 and December 2024. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study protocol was approved by the Ethics Committee of The Second Affiliated Hospital of Fujian Medical University {approval No. [2019]-89(203)} and the Ethics Committee of Jinjiang Municipal Hospital (approval No. jjsyyll-2026-066). The requirement for informed consent was waived because of the retrospective design.

Patients were eligible if they had pathologically confirmed primary breast cancer with ALN metastasis, underwent preoperative breast ultrasound before surgery, had available postoperative pathological assessment of ALNs, and had complete clinical and ultrasound data. Patients were excluded if they had a previous history of breast cancer, had received neoadjuvant chemotherapy or radiotherapy before surgery, had incomplete pathological or clinical data, or had ultrasound images of insufficient quality for radiomics analysis. After applying these criteria, 404 patients were included, with 294 patients in the training cohort and 110 patients in the external validation cohort. The patient selection process is shown in Figure 1.

Figure 1 Flowchart of patient selection for the training and validation cohorts. Patients were screened based on predefined inclusion and exclusion criteria. The final study population consisted of 404 patients, including 294 patients in the training cohort and 110 patients in the validation cohort.

ALN burden was defined according to the number of metastatic ALNs identified on postoperative pathological examination and was categorized as low nodal burden (1–2 metastatic lymph nodes) or high nodal burden (≥3 metastatic lymph nodes) according to criteria used in previous clinical trials and contemporary guidelines (17-19). Ultrasound-reported ALN status, cortical thickening, and hilum status were assessed based on the most suspicious ipsilateral ALN identified on preoperative ultrasound.

Clinical and conventional ultrasound variables collected for analysis included age, menopausal status, family history, tumor size, tumor side, tumor location, multifocality, lesion shape, lesion margin, posterior features, calcification, vascularity, ultrasound-reported ALN status, cortical thickening, hilum status, estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), Ki-67 proliferation index, molecular subtype, and axillary nodal burden. The sample size was determined by the availability of consecutive eligible patients during the study period at the two participating centers; no formal a priori sample size calculation was performed because of the retrospective study design.

Ultrasound radiomics analysis

Preoperative breast ultrasound examinations were performed as part of routine clinical practice using multiple commercially available ultrasound systems equipped with high-frequency linear-array transducers. At center A, the ultrasound systems included Mindray Resona 7OB, Mindray Resona I9, GE LOGIQ S7, GE LOGIQ Fortis, GE LOGIQ E9, GE Voluson E8, GE Voluson E10, and Siemens ACUSON Redwood. At center B, the ultrasound systems included GE LOGIQ E11, Philips EPIQ Elite, Mindray Resona R9, Hitachi ARIETTA 70, and Philips EPIQ 5G. The transducer frequency generally ranged from 7.5 to 15 MHz, depending on the equipment and lesion depth. For radiomics analysis, one representative grayscale image showing the largest tumor section with a clearly visible lesion boundary was selected for each patient. Images with severe artifacts, incomplete lesion display, or poor visualization of the tumor margin were excluded. Image selection and region of interest (ROI) delineation were performed by radiologists who were blinded to pathological axillary nodal burden. To reduce the influence of acquisition-related variability, reproducibility filtering based on intraclass correlation coefficients (ICCs) was performed, and feature standardization was fitted only in the training cohort and then applied unchanged to the external validation cohort.

Intratumoral and peritumoral ultrasound ROIs were analyzed. Tumor ROIs were manually delineated on ultrasound images using 3D Slicer software (version 4.11). Two radiologists independently performed tumor segmentation while being blinded to the pathological results to assess interobserver reproducibility.

For peritumoral radiomics analysis, a 3-mm peritumoral region was automatically generated outside the tumor boundary based on the original tumor contour. Radiomics features were subsequently extracted from both the intratumoral and peritumoral ROIs using the PyRadiomics package (version 3.0.1) implemented in Python (Figure 2).

Figure 2 Representative ultrasound images showing intratumoral and peritumoral ROI segmentation. (A) Original ultrasound image of the breast tumor. (B) Intratumoral ROI manually delineated along the tumor boundary. (C) Peritumoral ROI automatically generated as a 3-mm expansion outside the tumor contour. ROI, region of interest.

The extracted radiomics features included two-dimensional shape features, first-order statistical features, and texture features derived from gray-level matrices, including the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size-zone matrix (GLSZM), gray-level dependence matrix (GLDM), and neighboring gray-tone difference matrix (NGTDM). Both original and wavelet-transformed images were used for feature extraction.

During feature extraction, image normalization was applied with a scale of 100. Images were resampled to an in-plane resolution of 0.2×0.2 mm using B-spline interpolation, and gray-level discretization was performed using a fixed bin width of 5. Feature extraction was conducted in two-dimensional mode (force2D = true) to match the ultrasound acquisition characteristics.

To evaluate feature reproducibility, ICC analysis was performed, and features with ICC values >0.80 were retained for further analysis. After ICC filtering, 392 intratumoral features and 348 peritumoral features were preserved.

To avoid information leakage during external validation, whole-dataset ComBat harmonization was not used in the primary revised analysis. The training cohort and external validation cohort were separated before any preprocessing or model development. Feature standardization was fitted only in the training cohort and then applied unchanged to the validation cohort. An L1-penalized logistic regression model using inverse-frequency class weights was tuned over inverse regularization strengths (C values) from 10⁻⁴ to 1 by 5-fold cross-validation in the training cohort, with mean AUC as the selection criterion. Among features with nonzero coefficients at the selected C, the five features with the largest absolute coefficients were retained. A final class-weighted logistic regression model was then refitted using these five features.

Five features were ultimately selected for the intratumoral radiomics model and five for the peritumoral radiomics model. The selected intratumoral features included original_ngtdm_Complexity, wavelet-LL_firstorder_InterquartileRange, wavelet-LH_glcm_Idmn, original_ngtdm_Coarseness, and wavelet-LH_glszm_SizeZoneNonUniformityNormalized. The selected peritumoral features included wavelet-HL_firstorder_Maximum, original_ngtdm_Coarseness, wavelet-HH_glcm_Imc1, wavelet-LL_glcm_Imc2, and wavelet-LL_gldm_LargeDependenceLowGrayLevelEmphasis.

Clinical model development and integrated modeling

Clinical candidate variables were first screened in the training cohort. Univariate logistic regression analysis was performed for all candidate variables, and variables with P<0.10 in univariate analysis were subsequently entered into a multivariable logistic regression model.

Cortical thickening, hilum status, and ultrasound-reported ALN status were identified as independent clinical predictors of high axillary nodal burden and were used to construct the clinical model.

Six prespecified models were developed and compared: the clinical model, intratumoral radiomics model, peritumoral radiomics model, clinical + intratumoral model, clinical + peritumoral model, and clinical + intratumoral + peritumoral model. To reduce optimism in the combined models, out-of-fold radiomics scores from the training cohort were used for model integration. The external validation cohort was used only for independent performance evaluation and was not used for feature selection, threshold selection, parameter tuning, or model selection.

Statistical analysis

Continuous variables were expressed as mean ± standard deviation or median (interquartile range), as appropriate, whereas categorical variables were presented as counts and percentages. Normality was assessed using the Shapiro-Wilk test. Comparisons between cohorts were performed using the Student’s t-test or Mann-Whitney U test for continuous variables and the chi-square test or Fisher’s exact test for categorical variables.

Univariate logistic regression analysis was conducted to screen candidate clinical predictors, and variables with P<0.10 in univariate analysis were entered into multivariable logistic regression analysis. Model discrimination was evaluated using receiver operating characteristic (ROC) curves and the AUC, with 95% confidence intervals (CIs). Sensitivity, specificity, and accuracy were calculated using thresholds determined in the training cohort and then applied unchanged to the external validation cohort. Pairwise AUC comparisons in the external validation cohort were performed using bootstrap resampling.

Calibration was assessed using calibration curves, Brier score, calibration intercept, and calibration slope. DCA was used to evaluate potential net benefit across threshold probabilities. No model updating or recalibration was performed in the external validation cohort. A two-sided P value <0.05 was considered statistically significant.

The full model specifications, including selected features or predictors, coefficients, intercepts, and training-cohort standardization parameters, are provided in Tables S1-S3. For individual prediction, each feature was standardized as (raw value − training mean)/training standard deviation, and the predicted probability was calculated as 1/[1 + exp(−logit)].

Results

Baseline characteristics

A total of 404 patients were included in this study, comprising 294 patients in the training cohort and 110 patients in the validation cohort. The baseline clinicopathological and ultrasound characteristics of the two cohorts are summarized in Table 1. Most variables were comparable between the two cohorts. Age showed a borderline difference between the training and validation cohorts, whereas Ki-67 differed significantly between cohorts. In addition, the proportion of patients with high ALN burden was lower in the validation cohort than in the training cohort, although this difference did not reach statistical significance. Overall, the baseline comparability between the two cohorts was considered acceptable for subsequent model development and validation.

Clinical model construction

Univariate logistic regression identified cortical thickening, hilum status, ultrasound-reported ALN status, ER, PR, molecular subtype, tumor location, and Ki-67 as candidate predictors. In multivariable logistic regression, cortical thickening [odds ratio (OR) =4.384, P<0.001], hilum status (OR =2.455, P=0.003), and ultrasound-reported ALN status (OR =2.185, P=0.008) remained independent predictors of high axillary nodal burden. ER was not retained as an independent predictor in the revised multivariable model. The univariate and multivariable logistic regression results are shown in Table 2.

Radiomics model development

After ICC filtering and leakage-free standardization based on the training cohort, separate radiomics signatures were constructed for the intratumoral and peritumoral regions. The intratumoral radiomics model yielded an AUC of 0.632 in the training cohort and 0.621 (95% CI: 0.489–0.757) in the external validation cohort. The peritumoral radiomics model yielded AUCs of 0.689 and 0.585 (95% CI: 0.428–0.742), respectively. The feature selection process using LASSO regression is illustrated in Figure 3, and the full radiomics model specifications are provided in Tables S1,S2.

Figure 3 Illustration of the regularization paths for intratumoral and peritumoral radiomics features. (A,C) Cross-validated AUC profiles. (B,D) Coefficient profiles across the regularization path. The vertical reference lines indicate candidate values along the plotted paths. In the final analysis, the inverse regularization strength (C) was selected by 5-fold cross-validation using mean AUC; among features with nonzero coefficients, the five with the largest absolute magnitudes were retained, and the final radiomics models were refitted using these features. AUC, area under the curve; LASSO, least absolute shrinkage and selection operator.

The predictive performance of the six prespecified models is summarized in Table 3 and Figure 4. In the training cohort, the clinical + intratumoral and clinical + intratumoral + peritumoral models yielded the highest AUCs (both approximately 0.756), followed by the clinical model (AUC =0.740), clinical + peritumoral model (AUC =0.738), peritumoral model (AUC =0.689), and intratumoral model (AUC =0.632). In the external validation cohort, all models showed limited discrimination, with AUCs ranging from 0.585 to 0.621. The intratumoral radiomics model showed the numerically highest AUC (0.621), followed by the combined models (AUC =0.613), the clinical model (AUC =0.597), and the peritumoral model (AUC =0.585). Bootstrap pairwise comparisons showed no statistically significant differences between the intratumoral model and the other models (all P>0.05; Table S4).

Figure 4 ROC curves of six prespecified models for predicting high axillary nodal burden in the training and external validation cohorts. (A) ROC curves in the training cohort. The clinical + intratumoral and clinical + intratumoral + peritumoral models showed the highest apparent AUCs (both approximately 0.756), followed by the clinical model (AUC =0.740), clinical + peritumoral model (AUC =0.738), peritumoral radiomics model (AUC =0.689), and intratumoral radiomics model (AUC =0.632). (B) ROC curves in the external validation cohort. All models showed limited discrimination, with AUCs ranging from 0.585 to 0.621. The intratumoral radiomics model showed the numerically highest AUC (0.621), followed by the combined models (AUC =0.613), clinical model (AUC =0.597), and peritumoral radiomics model (AUC =0.585). The dashed diagonal line represents random classification. ALN, axillary lymph node; AUC, area under the curve; ROC, receiver operating characteristic.

Calibration metrics are summarized in Table 4. In the external validation cohort, Brier scores ranged from 0.235 to 0.255, and calibration slopes ranged from 0.042 to 0.827, indicating suboptimal calibration. DCA showed limited net benefit across the evaluated threshold range (Figure 5). The distribution of centered predicted probabilities from the intratumoral radiomics model is shown in Figure 6. These results suggest that the current models should be interpreted as exploratory risk stratification tools rather than clinically ready decision-making models.

Figure 5 Calibration curves and decision curve analysis of six prespecified models in the training and external validation cohorts. Calibration curves are shown for the training cohort (A) and external validation cohort (B). Decision curve analysis is shown for the training cohort (C) and external validation cohort (D). Calibration remained suboptimal in the external validation cohort, and decision curve analysis showed limited net benefit across the evaluated threshold range. DCA, decision curve analysis.

Figure 6 Waterfall plots of centered predicted probabilities from the intratumoral radiomics model in the training cohort (A) and external validation cohort (B). Patients are ranked in descending order. Bar colors indicate low and high axillary nodal burden.

Discussion

In this dual-center study, we revised the radiomics analysis using a leakage-free external validation workflow and compared six prespecified models for predicting high axillary nodal burden in node-positive breast cancer. Three main findings emerged. First, all models showed limited external discrimination, with validation AUCs ranging from 0.585 to 0.621. Second, the intratumoral radiomics model achieved the numerically highest validation AUC, but its advantage over the clinical, peritumoral, and combined models was not statistically significant in bootstrap comparisons. Third, although combined models showed relatively higher apparent performance in the training cohort, they did not improve external validation performance, and calibration remained suboptimal. These findings suggest that ultrasound radiomics may provide exploratory imaging information for nodal burden assessment, but the current models are not sufficient for standalone clinical decision-making.

Our findings are broadly consistent with the growing literature suggesting that ultrasound radiomics can assist in axillary assessment, but they also add a more cautious perspective regarding external robustness (6-12,19-21). As summarized in the Introduction, recent studies have reported promising results for ultrasound radiomics in predicting ALN status or nodal burden, and some of them emphasized the potential benefit of integrating intratumoral, peritumoral (8,9,15,20,21), and clinical information. However, much of the existing evidence is derived from retrospective, single-center cohorts with limited sample sizes, and relies predominantly on internal data splitting rather than independent external validation. In addition, most studies focus on predicting nodal metastasis status (i.e., presence or absence), whereas relatively few attempt to quantify nodal burden (9,16,19,22,23). Compared with many previous studies focusing on the presence of ALN metastasis, our study addressed a more challenging endpoint, namely the extent of ALN burden. Predicting nodal burden is inherently more difficult than predicting nodal positivity, as it requires finer discrimination of disease extent rather than the detection of a binary condition. In addition, many previously reported radiomics models were developed and validated within single-center datasets, which may lead to optimistic performance estimates. In contrast, the present study used a dual-center validation design, which provides a more stringent assessment of model generalizability.

A particularly important observation was the inconsistency between training and validation performance across different models. In the training cohort, combined models consistently achieved higher AUCs, which may reflect their ability to capture complementary information from multiple feature sources. However, this apparent advantage was not maintained in the validation cohort, although the intratumoral model showed the numerically highest validation AUC. This pattern suggests that the improved performance of combined models in the training cohort may be driven, at least in part, by overfitting. In radiomics studies with moderate sample sizes, integrating multiple feature domains increases model complexity and may amplify dataset-specific noise. Our findings therefore highlight the critical importance of external validation and suggest that the current findings do not support a clear advantage of more complex combined models in real-world settings.

Another noteworthy finding is the limited generalizability of the clinical model. Although the clinical model achieved a relatively high AUC in the training cohort (0.740), its performance decreased substantially in the validation cohort (0.597). This decline suggests that conventional clinical and ultrasound variables, even when statistically significant, may be sensitive to interobserver variability, institutional differences in imaging interpretation, and population heterogeneity. In contrast, the intratumoral radiomics model showed comparatively better external performance, indicating that quantitative imaging features may contain complementary information, although their external performance remained limited. This observation supports the potential role of radiomics as a tool to enhance robustness in multicenter applications, particularly for tasks such as ALN burden prediction where subtle imaging patterns may not be reliably assessed by visual evaluation alone.

Our findings also provide further insight into the ongoing debate regarding the value of peritumoral radiomics. The peritumoral region has been proposed as a surrogate for tumor-microenvironment interactions, including stromal response, vascular changes, and edema (16,24,25). While previous studies have reported improved performance when peritumoral features were incorporated, our results showed that the 3-mm peritumoral model underperformed the intratumoral model in the validation cohort, and that adding peritumoral features did not improve external discrimination (8,15,16,25,26). This discrepancy may be explained by several factors. First, peritumoral features are inherently more sensitive to ROI definition and surrounding tissue heterogeneity, which may reduce reproducibility across centers (13,14). Second, the optimal peritumoral width may be task- and modality-specific, and a fixed 3-mm expansion may not capture the most informative region for ultrasound-based ALN burden prediction. Third, compared with intratumoral features, peritumoral features may contain a higher proportion of background variability, limiting their incremental value in a moderate-sized dual-center dataset. Therefore, while peritumoral radiomics remains biologically appealing, its practical contribution should be interpreted with caution, particularly when external validation is considered.

From a clinical perspective, the present findings indicate that none of the evaluated models can currently be used as a standalone tool for axillary management. Although the intratumoral radiomics model showed the numerically highest AUC in the external validation cohort, its discrimination remained limited and was not statistically superior to the other models. In addition, calibration was suboptimal and DCA showed only limited net benefit. Therefore, the current models should be interpreted as exploratory tools for risk stratification rather than clinically ready decision-support systems. Future studies should define clinically acceptable thresholds, particularly with respect to sensitivity and false-negative rates, before such models can be considered for guiding axillary surgical decisions.

Several limitations should be acknowledged. First, this was a retrospective study with a moderate sample size, particularly in the external validation cohort, which included only 21 high-burden events. Therefore, the validation estimates, especially calibration metrics and sensitivity/specificity, may be unstable and should be interpreted cautiously. Second, only a single peritumoral width was evaluated, which limits conclusions regarding the full potential of peritumoral radiomics. Third, although a dual-center design was used, both cohorts were derived from a similar regional context, and broader multicenter validation is required. Finally, we used a conventional radiomics framework, and did not explore alternative modeling strategies, multimodal integration, or standardized acquisition protocols that may further improve robustness. Future studies should therefore include larger multicenter datasets, systematic evaluation of peritumoral regions, and model optimization strategies aimed at improving both generalizability and clinical utility. In addition, ultrasound images were retrospectively collected from routine clinical practice using multiple ultrasound systems across the two centers. Although this design reflects real-world imaging conditions and external validation was performed to assess generalizability, residual acquisition-related heterogeneity may have influenced radiomics feature stability and model performance. Future prospective studies using standardized acquisition protocols and scanner-specific quality control are warranted.

Conclusions

This study provides a conservative external validation assessment of ultrasound radiomics for predicting high axillary nodal burden in node-positive breast cancer. Under a leakage-free workflow, the predictive performance of all models was limited, and no model showed clear statistical superiority in the external validation cohort. The incremental value of 3-mm peritumoral radiomics and combined clinical-radiomics modeling was not confirmed. Further large-scale multicenter studies with standardized ultrasound acquisition, larger numbers of high-burden events, and prospective validation are needed before such models can be considered for clinical implementation.

Acknowledgments

We thank all the patients and staff at The Second Affiliated Hospital of Fujian Medical University and Jinjiang Municipal Hospital for their contributions to this work.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0226/rc

Data Sharing Statement: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0226/dss

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0226/prf

Funding: This work was supported by the Quanzhou Science and Technology Plan Project (No. 2019C072R).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0226/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 protocol was approved by the Ethics Committee of The Second Affiliated Hospital of Fujian Medical University {approval No. [2019]-89(203)} and the Ethics Committee of Jinjiang Municipal Hospital (approval No. jjsyyll-2026-066). The requirement for informed consent was waived because of the retrospective design.

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 inci-dence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209-49. [Crossref] [PubMed]
3. Chung HL, Le-Petross HT, Leung JWT. Imaging Updates to Breast Cancer Lymph Node Management. Radiographics 2021;41:1283-99. [Crossref] [PubMed]
4. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiol-ogy 2016;278:563-77.
5. Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and per-sonalized medicine. Nat Rev Clin Oncol 2017;14:749-62. [Crossref] [PubMed]
6. Guerrisi A, Miseo L, Falcone I, et al. Quantitative ultrasound radiomics analysis to evaluate lymph nodes in patients with cancer: a systematic review. Ultraschall Med 2024;45:586-96. [Crossref] [PubMed]
7. Zheng X, Yao Z, Huang Y, et al. Deep learning radiomics can predict axillary lymph node status in early-stage breast cancer. Nat Commun 2020;11:1236. [Crossref] [PubMed]
8. Wang SR, Cao CL, Du TT, et al. Machine Learning Model for Predicting Axillary Lymph Node Me-tastasis in Clinically Node Positive Breast Cancer Based on Peritumoral Ultrasound Radiomics and SHAP Feature Analysis. J Ultrasound Med 2024;43:1611-25. [Crossref] [PubMed]
9. Wang SR, Tian F, Zhu T, et al. Machine learning-driven ultrasound radiomics for assessing axillary lymph node burden in breast cancer. Front Endocrinol (Lausanne) 2025;16:1548888. [Crossref] [PubMed]
10. Zhang W, Wang S, Wang Y, et al. Ultrasound-based radiomics nomogram for predicting axillary lymph node metastasis in early-stage breast cancer. Radiol Med 2024;129:211-21. [Crossref] [PubMed]
11. Liu H, Zou L, Xu N, et al. Deep learning radiomics based prediction of axillary lymph node metastasis in breast cancer. NPJ Breast Cancer 2024;10:22. [Crossref] [PubMed]
12. Liu Z, Hong M, Li X, et al. Predicting axillary lymph node metastasis in breast cancer patients: A ra-diomics-based multicenter approach with interpretability analysis. Eur J Radiol 2024;176:111522. [Crossref] [PubMed]
13. Park JE, Park SY, Kim HJ, et al. Reproducibility and Generalizability in Radiomics Modeling: Possible Strategies in Radiologic and Statistical Perspectives. Korean J Radiol 2019;20:1124-37. [Crossref] [PubMed]
14. Mali SA, Ibrahim A, Woodruff HC, et al. Making Radiomics More Reproducible across Scanner and Imaging Protocol Variations: A Review of Harmonization Methods. J Pers Med 2021;11:842. [Crossref] [PubMed]
15. Sun Q, Lin X, Zhao Y, et al. Deep Learning vs. Radiomics for Predicting Axillary Lymph Node Me-tastasis of Breast Cancer Using Ultrasound Images: Don't Forget the Peritumoral Region. Front Oncol 2020;10:53.
16. Luo H, Chen Z, Xu H, et al. Peritumoral edema enhances MRI-based deep learning radiomic model for axillary lymph node metastasis burden prediction in breast cancer. Sci Rep 2024;14:18900. [Crossref] [PubMed]
17. Giuliano AE, Hunt KK, Ballman KV, et al. Axillary dissection vs no axillary dissection in women with invasive breast cancer and sentinel node metastasis: a randomized clinical trial. JAMA 2011;305:569-75. [Crossref] [PubMed]
18. Giuliano AE, Ballman KV, McCall L, et al. Effect of Axillary Dissection vs No Axillary Dissection on 10-Year Overall Survival Among Women With Invasive Breast Cancer and Sentinel Node Metastasis: The ACOSOG Z0011 (Alliance) Randomized Clinical Trial. JAMA 2017;318:918-26. [Crossref] [PubMed]
19. Zhang X, Yang Z, Cui W, et al. Preoperative prediction of axillary sentinel lymph node burden with multiparametric MRI-based radiomics nomogram in early-stage breast cancer. Eur Radiol 2021;31:5924-39. [Crossref] [PubMed]
20. Ma Q, Wang J, Tu Z, et al. Prediction model of axillary lymph node status using an automated breast volume ultrasound radiomics nomogram in early breast cancer with negative axillary ultrasound. Front Immunol 2025;16:1460673. [Crossref] [PubMed]
21. Ye X, Zhang X, Lin Z, et al. Ultrasound-based radiomics nomogram for predicting axillary lymph node metastasis in invasive breast cancer. Am J Transl Res 2024;16:2398-410. [Crossref] [PubMed]
22. Yao J, Zhou W, Xu S, et al. Machine Learning-Based Breast Tumor Ultrasound Radiomics for Pre-operative Prediction of Axillary Sentinel Lymph Node Metastasis Burden in Early-Stage Invasive Breast Cancer. Ultrasound Med Biol 2024;50:229-36. [Crossref] [PubMed]
23. Xu J, Qi P, Ou X, et al. Bi-modal ultrasound radiomics and habitat analysis enhanced the pre-operative prediction of axillary lymph node burden in patients with early-stage breast cancer. Front Oncol 2025;15:1607442. [Crossref] [PubMed]
24. Cheon H, Kim HJ, Kim TH, et al. Invasive Breast Cancer: Prognostic Value of Peritumoral Edema Identified at Preoperative MR Imaging. Radiology 2018;287:68-75. [Crossref] [PubMed]
25. Park NJ, Jeong JY, Park JY, et al. Peritumoral edema in breast cancer at preoperative MRI: an interpre-tative study with histopathological review toward understanding tumor microenvironment. Sci Rep 2021;11:12992. [Crossref] [PubMed]
26. Yang W, Su X, Yue C, et al. Interpretable machine learning based on intratumoral and peritumoral ul-trasound radiomics for predicting central lymph node metastasis in papillary thyroid carcinoma. Eur J Radiol 2026;197:112727. [Crossref] [PubMed]