Abstract
Background
To determine the best-performing peritumoral boundary among tested distances and evaluate the predictive value of preoperative multiparametric MRI radiomics for axillary lymph node metastasis (ALNM) in breast cancer.
Methods
Clinical and imaging data of 273 patients from two centers were retrospectively analyzed. ROIs were manually delineated on Fs-T2WI and DCE-MRI images, and peritumoral regions were generated by expanding the tumor ROI by 3 mm, 5 mm, and 7 mm. Patients from Center 1 (n = 201) were randomly divided into training and internal validation set (7:3), while Center 2 (n = 72) served as the external validation set. LASSO regression and an SVM classifier were used to construct predictive models. A nomogram integrating the best-performing radiomic model with clinical parameters was developed. Model performance was assessed using AUC and DCA.
Results
Maximum tumor diameter was identified as an independent risk factor for ALNM (OR = 4.042, 95% CI: 2.488–6.532; P < 0.001). Among the tested peritumoral margins, the 5 mm model demonstrated the best predictive efficacy, significantly outperforming both 3 mm (AUC 0.694) and 7 mm (AUC 0.717) models (DeLong test, all P < 0.05). The fusion model combining DCE-MRI intra- and peritumoral (5 mm) features with Fs-T2WI intratumoral features achieved AUCs of 0.909, 0.818, and 0.768 in the training, internal validation, and external validation sets, respectively. The nomogram integrating the fusion model with maximum tumor diameter further improved performance, with AUCs of 0.943, 0.884, and 0.853, respectively.
Conclusion
Among the tested distances, a 5 mm peritumoral margin provided the best predictive performance for ALNM. The multiparametric MRI radiomic model effectively predicts ALNM and offers a valuable non-invasive tool for preoperative assessment.
Keywords: radiomics, breast cancer, magnetic resonance imaging, peritumoral region, axillary lymph node metastasis
Introduction
Breast cancer is one of the most prevalent malignancies in women worldwide and is the leading cause of cancer-related mortality.1 Axillary lymph node (ALN) status is critical for tumor staging, therapeutic decision-making, and prognostic assessment in breast cancer. It is well established that the 10 year survival rate exceeds 90% in patients without nodal metastasis, whereas it declines substantially in those with positive nodes.2 For patients with early-stage breast cancer who are clinically ALN-negative or have only a few positive sentinel lymph nodes, complete axillary lymph node dissection can safely be omitted. Nevertheless, sentinel lymph node biopsy is an invasive procedure with inherent risks, including lymphedema, upper limb pain, and functional impairment of the upper extremity.3 Therefore, developing a noninvasive, reliable, preoperative evaluation approach holds considerable clinical significance.
Radiomics transforms medical images into high-dimensional quantitative data, enabling the noninvasive characterization of tumor heterogeneity and the peritumoral microenvironment. This information largely remains imperceptible to visual inspection. This approach has gained considerable traction in oncological imaging and has emerging applications in predicting prognosis, treatment response, and molecular subtyping.4 In recent years, radiomics has been extensively investigated across various tumor types. Within the field of breast oncology, mounting evidence suggests that radiomic models derived from magnetic resonance imaging (MRI) show great promise for tumor detection, differential diagnosis, and predicting axillary lymph node metastasis (ALNM).5–7
Previous radiomic investigations have primarily focused on extracting and analyzing intratumoral features. However, a growing body of evidence suggests that the tumor microenvironment plays a key role in regulating angiogenesis, immune modulation, and metastatic progression in breast cancer.8,9 The peritumoral region contains critical pathological information, including lymphovascular invasion, inflammatory cell infiltration, and desmoplastic reaction. All of these can be quantified through radiomic feature extraction. Magnetic resonance imaging (MRI), with its superior soft tissue contrast and multiparametric acquisition capability, provides additional biological insights. Dynamic contrast-enhanced (DCE-MRI) sequences can assess tumor blood perfusion and vascular permeability, and fat-suppressed T2-weighted (Fs-T2WI) sequences can detect peritumoral edema and stromal reaction.10 Emerging evidence indicates that MRI-derived peritumoral radiomic features are closely correlated with tumor biological behavior. Models that integrate both intratumoral and peritumoral features have been shown to consistently outperform those that only consider intratumoral features in predicting lymph node metastasis.11
However, the optimal peritumoral boundary for radiomic analysis remains controversial, because previous studies have used widely varying expansion distances across different imaging modalities. In MRI-based radiomics, Wu et al12 systematically compared multiple peritumoral margins and identified a 4 mm region as optimal for predicting ALNM, whereas other investigators reported that a 3 mm margin yielded superior performance.13 In ultrasound imaging, Duan et al14 evaluated seven expansion distances ranging from 0.5 to 3.5 mm and found that the 0.5 mm peritumoral region performed best for ALNM prediction. Meanwhile, a digital breast tomosynthesis (DBT) based study demonstrated that combining intratumoral features with a 10 mm peritumoral region achieved the highest predictive performance.15 Beyond ALNM prediction, Huang et al16 extracted radiomic features from photoacoustic imaging at peritumoral distances of 2, 5, and 8 mm and found that a 5 mm margin most effectively discriminated between benign and malignant breast nodules, and Jiang et al17 reported that a peritumoral MRI radiomic model effectively predicted lymphovascular invasion. These discrepancies across studies may be attributable to differences in spatial resolution across imaging modalities, heterogeneity of the study populations, and variation in the ranges of expansion distances evaluated.
Against this background, we systematically compared three peritumoral expansion distances (3, 5, and 7 mm) on DCE-MRI. The 5 mm margin was selected as the reference because of its widespread adoption in previous radiomic studies,18,19 while the 3 mm and 7 mm margins were added to determine whether narrower or wider peritumoral regions could further improve ALNM prediction. Given that multiparametric MRI can provide complementary biological information, integrating DCE-MRI derived intra and peritumoral features with Fs-T2WI derived intratumoral features may further enhance predictive performance. We accordingly constructed a multiparametric radiomic model that incorporated features from both sequences, and subsequently developed a clinically applicable nomogram combining the best-performing radiomic signature with independent clinical risk factors to facilitate individualized preoperative risk stratification.
Materials and Methods
Patients
The study cohort consisted of female patients with pathologically confirmed invasive breast cancer who underwent breast magnetic resonance imaging (MRI) examinations at our institution (Center 1) between May 2022 and December 2025, as well as at another large, comprehensive hospital (Center 2) between September 2025 and January 2026. Patients from Center 1 were randomly assigned to the training and internal validation sets at a ratio of 7:3. Patients from Center 2 constituted the external validation set. This retrospective study was approved by the Institutional Ethics Committee of the First Hospital of Anhui University of Science and Technology (Approval No. 2025KYY039001).
The inclusion criteria were as follows: (1) histopathologically confirmed invasive breast carcinoma, (2) underwent axillary lymph node dissection or sentinel lymph node biopsy, and (3) had no history of preoperative radiotherapy, chemotherapy, neoadjuvant therapy, or other nonsurgical interventions. The exclusion criteria were: (1) incomplete clinical, imaging, or pathological data, (2) poor image quality that precludes accurate lesion delineation, and (3) prior breast cancer surgery or chemotherapy. Figure 1 illustrates the detailed study flowchart and design.
Figure 1.
Flowchart of the patients included in this study.
Clinicopathological Information
Baseline clinical and pathological data, including age, estrogen receptor (ER) status, progesterone receptor (PR) status, Ki-67 proliferation index, histological grade, lymphovascular invasion (LVI) status, and axillary sentinel lymph node metastasis status, were retrieved from electronic medical records. Based on pathological assessment, axillary lymph node metastasis was confirmed in 138 patients, while the remaining 135 patients were node negative.
Image Acquisition and Preprocessing
All MRI examinations were performed using a GE Signa Discovery MR 750W 3.0T scanner (GE Healthcare, Milwaukee, WI, USA) equipped with an 8, channel dedicated breast coil. All patients were scanned in the prone position and underwent axial fat-suppressed T2-weighted imaging (Fs-T2WI), axial non-fat-suppressed T1-weighted imaging (T1WI), diffusion-weighted imaging (DWI) with apparent diffusion coefficient (ADC) mapping, and dynamic contrast-enhanced MRI (DCE-MRI). Gadopentetate dimeglumine was intravenously administered at a dose of 0.2 mmol/kg at a flow rate of 2 mL/s. One pre-contrast phase was acquired as a mask, followed by seven post-contrast phases. The third post-contrast phase (approximately 180 seconds after contrast injection) was selected for tumor segmentation, as it provided optimal tumor enhancement and clearly delineated lesion boundaries.11
Image Segmentation and Feature Extraction
To ensure the reliability and reproducibility of subsequent radiomic analyses, all images underwent standardized preprocessing before ROI segmentation. The preprocessing pipeline included the following steps: (1) resampling to an isotropic voxel size of 1 × 1×1 mm3 using cubic spline interpolation to minimize variability arising from differences in acquisition parameters, (2) N4 bias field correction to mitigate the effects of B0 field inhomogeneity, and (3) score normalization to standardize gray level intensity ranges across different participants and imaging platforms.
Region of interest (ROI) delineation was performed as follows: The required MRI sequences (DCE-MRI third phase and Fs-T2WI) were exported from the PACS workstation in DICOM format and imported into ITK-SNAP software (version 3.8; https://www.itksnap.org). Two radiologists, each with over five years of experience in breast imaging diagnosis, independently delineated the breast tumors in a blinded manner (without knowledge of pathological results). On both DCE-MRI and Fs-T2WI images, three-dimensional ROIs were manually segmented layer by layer along the tumor margin (Figure 2).
Figure 2.
Examples of delineating regions of interest (ROIs) within and around tumors using ITK-SNAP software. (a–e) DCE image, intratumoral ROI, and peritumoral ROI; (f–g) Fs-T2WI image and intratumoral ROI. Red indicates intratumoral ROI, and green indicates peritumoral ROI.
The delineation process strictly adhered to the following principles: (1) Segmentation was confined to the lesion margins while incorporating necrotic and cystic components to account for intratumoral heterogeneity. (2) In cases of disagreement between the two radiologists, a consensus was reached through discussion or adjudication by a senior radiologist with over ten years of experience. (3) For patients with multiple breast lesions, only the largest lesion (by maximum diameter) was included in the analysis.
Peritumoral Margin Definition and Application
To identify the optimal peritumoral margin for radiomic analysis, the intratumoral region of interest (ROI) on dynamic contrast-enhanced (DCE) MRI images was automatically expanded outward using ITK-SNAP software at three predefined distances: 3 mm, 5 mm, and 7 mm. Specifically, three peritumoral regions—denoted as ROI_Peri3mm, ROI_Peri5mm, and ROI_Peri7mm—were generated by expanding the original tumor ROI along the boundary; the intratumoral region was excluded from each peritumoral volume. Throughout the entire delineation process, the two radiologists remained blinded to all clinical and pathological information.
Based on the preliminary comparative analysis, identified as the best-performing peritumoral margin among the tested distances and was adopted for DCE-MRI in subsequent model construction. For Fs-T2WI, only intratumoral features were utilized because the peritumoral margin for this sequence was not optimized independently in the present study.
Radiomics Feature Extraction
Following ROI segmentation and image preprocessing, radiomic features were extracted from the intratumoral and peritumoral (5 mm) regions of DCE-MRI and from the intratumoral region of Fs-T2WI using the PyRadiomics package (version 3.0; https://pypi.org/project/pyradiomics/3.0/) implemented in Python (version 3.7). The extracted feature set comprised first-order statistics, shape-based features, and multiple texture feature classes, 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).
To ensure the robustness and reproducibility of the extracted features, a reproducibility analysis was performed on a randomly selected subset of 42 patients. Two radiologists independently delineated the ROIs, and the procedure was repeated by one of the radiologists after a one-week interval to assess intra-observer agreement. The intraclass correlation coefficient (ICC) was calculated to evaluate both inter- and intra-observer reliability. Features with an ICC ≥ 0.75 were considered reproducible and retained for subsequent analyses, whereas those with an ICC < 0.75 were excluded to minimize variability attributable to segmentation discrepancies.
Radiomics Feature Selection and Model Construction
Feature Selection and Data Normalization
Following feature extraction, a multi-step feature selection strategy was implemented to identify the most discriminative predictors and mitigate the risk of overfitting. First, features exhibiting near-zero variance (variance < 0.01) were excluded. Second, Spearman correlation analysis was performed to remove highly correlated features with a correlation coefficient exceeding 0.9 to reduce redundancy. Subsequently, the least absolute shrinkage and selection operator (LASSO) logistic regression algorithm, incorporating 5-fold cross-validation, was applied to select the optimal feature subset, with the regularization parameter λ determined using the 1-standard error criterion.
All features were normalized to a mean of 0 and standard deviation of 1 using Z-score standardization before model construction to eliminate the influence of different feature scales. The normalization formula used was: Xnorm=(X−μ)/σ, where μ and σ represent the mean and standard deviation of each feature calculated from the training set; the same μ and σ values were subsequently applied to normalize the internal validation and external validation sets.
Radiomics Model Construction
Radiomics scores (Rad-scores) were calculated for each ROI as the linear combination of the selected features weighted by their respective LASSO coefficients. Multiple radiomic models were subsequently constructed using a support vector machine (SVM) classifier with a radial basis function (RBF) kernel. The SVM hyperparameters, including the cost parameter C and the kernel coefficient gamma, were optimized by grid search over C∈{0.1, 1, 10, 100} and gamma∈{0.01, 0.1, 1, 10}, and the combination yielding the highest average AUC across five-fold cross-validation on the training set was selected for the final models.
The following models were developed in two phases: (1) Phase 1—single-region peritumoral models (DCE-peri3mm, DCE-peri5mm, and DCE-peri7mm) to identify the best-performing peritumoral margin among the tested distances; and (2) Phase 2—comprehensive models incorporating the selected DCE-MRI peritumoral margin in combination with Fs-T2WI features.
Clinical Model and Nomogram
Univariate and multivariate logistic regression analyses were performed to identify independent clinical risk factors for axillary lymph node (ALN) metastasis. A clinical model was then constructed based on the identified predictors. Finally, a nomogram was developed by integrating the best-performing radiomic model with the independent clinical predictors using the “rms” package in R software (version 4.2; https://www.r-project.org/).
Model Evaluation
Model performance was evaluated in the training (70%), internal validation set (30%), and external validation (Center 2) sets using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Pairwise comparisons of AUCs among the radiomic models, as well as between the nomogram and alternative models, were performed using the DeLong test, and the 95% confidence intervals (CIs) of the AUCs were estimated using the bootstrap method with 1000 resampling iterations. Model calibration was evaluated using calibration curves and the Hosmer-Lemeshow goodness-of-fit test, and clinical utility was assessed using decision curve analysis (DCA). The overall workflow of radiomic analysis is depicted in Figure 3.
Figure 3.
Flowchart of multi-parameter MRI radiomic analysis.
Statistical Analysis
Radiomic feature extraction, model construction, and statistical comparisons were performed using R software (version 4.2; https://www.r-project.org/) and Python (version 3.7; https://www.python.org/), and clinical and pathological data analyses were performed using SPSS software (version 27.0; IBM Corp., Armonk, NY, USA). Continuous variables were first assessed for normality and homogeneity of variance. Normally distributed variables were expressed as mean ± standard deviation and compared between groups using the independent-samples t test, whereas non-normally distributed variables were expressed as the median with interquartile range (M [Q1, Q3]) and compared using the Mann–Whitney U-test. Categorical variables were compared using the Pearson χ2-test or Fisher’s exact test, as appropriate. All statistical tests were two-sided, and P < 0.05 was considered statistically significant.
Results
Patient Characteristics and Clinical Model
A total of 273 patients from two independent centers were enrolled in this study. The baseline clinicopathological characteristics of patients from Center 1 (n = 201) and Center 2 (n = 72) are summarized in Table 1. Baseline characteristics including age, ER/PR status, Ki-67 index, and enhancement patterns were comparable between ALNM-positive and ALNM-negative cohorts (all P > 0.05). However, maximum tumor diameter and spiculation differed significantly between the two groups in both centers (all P < 0.05).
Table 1.
Clinicopathological Characteristics in Different Centers
| Parameters | Center 1, N = 201 | Center 2, N = 72 | ||||
|---|---|---|---|---|---|---|
| ALNM-Negative (n = 97) | ALNM-Positive (n = 104) | P value | ALNM-Negative (n = 38) | ALNM-Positive (n = 34) | P value | |
| Age | 55.39±10.15 | 56.04±10.94 | 0.665 | 56.86±9.66 | 57.74±11.06 | 0.712 |
| ER Status | 0.823 | |||||
| Without | 28 (28.9%) | 29 (27.9%) | 11 (28.9%) | 10 (29.4%) | 0.905 | |
| With | 69 (71.1%) | 75 (72.1%) | 27 (71.1%) | 24 (70.6%) | ||
| PR Status | 0.756 | 0.891 | ||||
| Without | 35 (36.1%) | 39 (37.5%) | 14 (36.8%) | 13 (38.2%) | ||
| With | 62 (63.9%) | 65 (62.5%) | 24 (63.2%) | 21 (61.8%) | ||
| Ki-67 Status | 0.542 | 0.678 | ||||
| ≥20% | 58 (59.8%) | 65 (62.5%) | 23 (60.5%) | 20 (58.8%) | ||
| ≤20% | 39 (40.2%) | 39 (37.5%) | 15 (39.5%) | 14 (41.2%) | ||
| Rim enhancement | 0.451 | 0.612 | ||||
| Without | 72 (74.2%) | 75 (72.1%) | 28 (73.7%) | 26 (76.5%) | ||
| With | 25 (25.8%) | 29 (27.9%) | 10 (26.3%) | 8 (23.5%) | ||
| ADC×10−3mm2/s | 1.12 ± 0.28 | 1.08 ± 0.31 | 0.642 | 1.15 ± 0.26 | 1.09 ± 0.29 | 0.401 |
| Maximum diameter | 21.5 ± 8.2 | 24.6 ± 10.5 | <0.001 | 20.2 ± 7.8 | 24.42± 11.2 | <0.001 |
| Spiculation | 0.032 | 0.041 | ||||
| Without | 71 (73.2%) | 62 (59.6%) | 28 (73.7%) | 20 (58.8%) | ||
| With | 26 (26.8%) | 42 (40.4%) | 10 (26.3%) | 14 (41.2%) | ||
| Enhanced pattern | 0.891 | 0.756 | ||||
| Mass | 82 (84.5%) | 87 (83.7%) | 32 (84.2%) | 28 (82.4%) | ||
| Non-mass | 15 (15.5%) | 17 (16.3%) | 6 (15.8%) | 6 (17.6%) | ||
Abbreviations: ER, Estrogen Receptor; PR, Progesterone Receptor; Ki67, Proliferation Marker Protein Ki67; ALN, Axillary Lymph Node.
After multivariable adjustment, only maximum tumor diameter remained independently associated with ALNM (OR 4.042, 95% CI 2.488–6.532; P < 0.001); spiculation lost significance upon controlling for this variable (P = 0.256). The clinical model incorporating maximum tumor diameter yielded area under the curve (AUC) values of 0.787 in the training set, 0.743 in the internal validation set, and 0.786 in the external validation set.
Peritumoral Margin Comparison
Among the single-region peritumoral models derived from DCE-MRI, the 5-mm margin consistently outperformed the 3 mm and 7 mm margins across both the training and internal validation sets. Specifically, the DCE-MRI peri3 mm model yielded area under the curve (AUC) values of 0.694 and 0.672 in the training and internal validation sets, respectively. Meanwhile, the DCE-MRI peri7 mm model achieved AUC values of 0.717 and 0.643, respectively (Table 2). The DeLong test confirmed that, in the training set, DCE-peri5 mm performed significantly better than DCE-peri3 mm (Z = 2.501, P = 0.012) and DCE-peri7 mm (Z = 2.820, P = 0.004). Consistent findings were observed in the test set, where the DCE peri5 mm model also showed significantly higher AUCs than the DCE peri3mm and DCE peri7 mm models (Z = 2.371, P = 0.024 and Z = 2.619, P = 0.005, respectively; Table 3). On the basis of these results, the 5 mm peritumoral margin was identified as the best-performing peritumoral margin among the tested distances and was subsequently adopted for multiparametric model construction.
Table 2.
Performance of Radiomics Models
| Radiomics Model | Training Set | Internal Validation Set | ||||
|---|---|---|---|---|---|---|
| AUC (95% CI) | Specificity | Sensitivity | AUC (95% CI) | Specificity | Sensitivity | |
| DCE_peri3mm | 0.694 (0.681–0.783) | 0.716 | 0.714 | 0.672 (0.542–0.756) | 0.715 | 0.734 |
| DCE_peri5mm | 0.754 (0.744–0.822) | 0.878 | 0.679 | 0.723 (0.640–0.837) | 0.663 | 0.803 |
| DCE_peri7mm | 0.717 (0.657–0.831) | 0.851 | 0.571 | 0.643 (0.523–0.725) | 0.845 | 0.745 |
Table 3.
DeLong Test for Comparing Peritumoral Radiomics Models
| Delong Test | Training Set | Internal Validation Set | ||
|---|---|---|---|---|
| Z value | P value | Z value | P value | |
| Peri 3mm vs Peri 5mm | 2.501 | 0.012 | 2.371 | 0.024 |
| Peri 3mm vs Peri 7mm | −0.701 | 0.748 | −0.605 | 0.516 |
| Peri 5mm vs Peri 7mm | 2.820 | 0.004 | 2.619 | 0.005 |
Multiparametric Radiomics Models
Feature Selection
Using the best-performing 5 mm peritumoral margin among the tested distances, LASSO regression was applied to select the most informative features for each model. The T2WI intra model retained six intratumoral features, and the DCE intra+peri5mm model retained 17 optimal features, comprising nine intratumoral and eight peritumoral features. The fusion model, which integrated features from both sequences, retained 20 optimal features, including 14 derived from DCE-MRI and six from Fs-T2WI.
Among the selected features in the fusion model, the most influential predictors with the largest absolute LASSO coefficients were wavelet-LLL_glcm_Correlation (coefficient = 0.418), log-sigma-3-0-mm_3D_ngtdm_Busyness (coefficient = 0.412), wavelet-HHL_glszm_LargeZoneLowGrayLevelEmphasis (coefficient = −0.358), wavelet-LLL_firstorder_Uniformity (coefficient = −0.345), and wavelet-HLH_glszm_GrayLevelNonUniformity (coefficient = 0.204). Notably, wavelet-transformed texture features predominated among the top predictors, suggesting that high-frequency textural information reflecting tumor heterogeneity and peritumoral microenvironment may be particularly informative for ALNM prediction.
Performance of Single-Sequence and Fusion Models
Among the DCE-MRI-based models, the intra model yielded AUCs of 0.780, 0.731, and 0.692 in the training, internal validation, and external validation sets, respectively. The peri5mm model achieved AUCs of 0.754, 0.723, and 0.663, respectively, while the intra+peri5mm model produced AUCs of 0.821, 0.773, and 0.732, respectively. The Fs-T2WI-intra model produced AUCs of 0.759, 0.702, and 0.674 across the three cohorts.
Notably, the fusion model, which combines DCE-MRI intra and peritumoral (5 mm) features with Fs-T2WI intratumoral features, exhibited the highest predictive performance among all radiomic models. It achieved AUCs of 0.909 (95% CI: 0.860–0.958) in the training set, 0.818 (95% CI: 0.740–0.914) in the internal validation set, and 0.768 (95% CI: 0.660–0.876) in the external validation set (Table 4).
Table 4.
Comparison of Diagnostic Efficacy of DCE, Fs-T2WI, Fusion, Clinical and Nomogram Models
| Model | Dataset | AUC | 95% CI | Sensitivity | Specificity | PPV | NPV |
|---|---|---|---|---|---|---|---|
| Fs-T2WI_intra | Training set | 0.759 | 0.683–0.822 | 0.724 | 0.808 | 0.802 | 0.734 |
| Internal validation set | 0.702 | 0.601–0.793 | 0.720 | 0.707 | 0.725 | 0.702 | |
| External validation set | 0.674 | 0.582–0.763 | 0.650 | 0.682 | 0.647 | 0.685 | |
| DCE_intra | Training set | 0.780 | 0.703–0.845 | 0.735 | 0.812 | 0.807 | 0.741 |
| Internal validation set | 0.731 | 0.641–0.817 | 0.742 | 0.731 | 0.746 | 0.725 | |
| External validation set | 0.692 | 0.612–0.796 | 0.684 | 0.711 | 0.679 | 0.715 | |
| DCE_intra+peri5mm | Training set | 0.821 | 0.763–0.879 | 0.765 | 0.726 | 0.750 | 0.742 |
| Internal validation set | 0.773 | 0.661–0.837 | 0.720 | 0.649 | 0.687 | 0.684 | |
| External validation set | 0.732 | 0.648–0.814 | 0.740 | 0.630 | 0.642 | 0.730 | |
| Fusion | Training set | 0.909 | 0.860–0.958 | 0.850 | 0.871 | 0.876 | 0.844 |
| Internal validation set | 0.818 | 0.740–0.914 | 0.801 | 0.780 | 0.796 | 0.785 | |
| External validation set | 0.768 | 0.660–0.876 | 0.760 | 0.720 | 0.708 | 0.770 | |
| Clinical | Training set | 0.787 | 0.690–0.854 | 0.710 | 0.748 | 0.751 | 0.706 |
| Internal validation set | 0.743 | 0.621–0.836 | 0.767 | 0.782 | 0.789 | 0.758 | |
| External validation set | 0.786 | 0.680–0.872 | 0.739 | 0.743 | 0.720 | 0.761 | |
| Nomogram | Training set | 0.943 | 0.910–0.976 | 0.901 | 0.917 | 0.921 | 0.896 |
| Internal validation set | 0.884 | 0.840–0.932 | 0.835 | 0.880 | 0.882 | 0.833 | |
| External validation set | 0.853 | 0.792–0.884 | 0.820 | 0.801 | 0.787 | 0.833 |
Nomogram Performance
The nomogram combining maximum tumor diameter with the fusion radiomic signature demonstrated excellent discriminatory performance, with AUCs of 0.943 (95% CI: 0.910–0.976), 0.884 (95% CI: 0.840–0.932), and 0.853 (95% CI: 0.792–0.884) in the training, internal validation, and external validation sets, respectively (Table 4 and Figure 4). DCA also showed superior net clinical benefit compared with the clinical model or radiomic signature alone (Figures 5 and 6).
Figure 4.
ROC curves of the clinical, fusion, and nomogram models for predicting axillary lymph node metastasis. (a) Training set; (b) Internal validation set; (c) External validation set. Different colors and line styles represent the clinical model, fusion model, and nomogram, respectively.
Figure 5.
Decision curve analysis of the clinical, fusion, and combined models for predicting axillary lymph node metastasis. (a) Training set; (b) Internal validation set; (c) External validation set. Different colors and line styles represent the clinical model, fusion model, and combined model, respectively.
Figure 6.
(a) LASSO coefficient profiles of selected features; (b) Nomogram for predicting axillary lymph node metastasis.
Discussion
This two-stage study yielded several important findings. First, a systematic comparison of 3 mm, 5 mm, and 7 mm peritumoral margins on DCE-MRI revealed that the 5 mm margin offered the best predictive performance for axillary lymph node metastasis (ALNM), outperforming the narrower and wider margins significantly. Second, the multiparametric MRI radiomic model integrating intra- and peritumoral (5 mm) features from DCE-MRI and intratumoral features from Fs-T2WI achieved excellent predictive efficacy, with area under the curve (AUC) values of 0.909 in the training set, 0.818 in the internal validation set, and 0.768 in the external validation set. Third, the nomogram combining the optimal radiomic signature with maximum tumor diameter exhibited outstanding calibration and clinical utility, attaining AUCs of 0.943 in the training set and 0.884 and 0.853 in the internal validation set and external validation set, respectively. Collectively, these findings underscore the potential of the proposed imaging-based approach for non-invasive preoperative risk stratification of ALNM in breast cancer patients.
Our initial findings suggest that the 5 mm peritumoral margin strikes a favorable balance by capturing sufficient microenvironmental information while minimizing signal contamination from adjacent normal breast tissue. This finding is consistent with previous reports showing that a 5 mm expansion effectively encompasses peritumoral alterations associated with tumor invasion and metastatic progression.17,20 In contrast, the inferior performance of the 3 mm margin may be largely attributable to its insufficient coverage of the peritumoral microenvironment. Specifically, a 3 mm expansion may fail to fully capture the extent of lymphatic vessel invasion, inflammatory infiltration, and desmoplastic reaction, all of which frequently extend beyond the immediate tumor boundary. These pathological changes are well-established determinants of metastatic potential, and incomplete representation of them is likely to compromise predictive accuracy. In support of this notion, Ding et al21 reported that a peritumoral radiomic model significantly outperformed an intratumoral-only counterpart in predicting lymphatic invasion, further underscoring the importance of adequate peritumoral sampling.
Conversely, the suboptimal performance observed at the 7 mm margin suggests that excessive expansion introduces noise from adjacent normal breast parenchyma, thereby diluting the signal-to-noise ratio of genuine peritumoral features. As the expansion distance increases, the proportion of normal tissue within the peritumoral region of interest (ROI) also increases, potentially incorporating irrelevant textural patterns that do not reflect authentic tumor-microenvironment interactions.14,22 This phenomenon is consistent with the “diminishing returns” principle documented in other radiomic studies, wherein overly broad peritumoral regions fail to improve—and may even degrade—predictive performance. Notably, our findings are in concordance with those of Huang et al, who combined photoacoustic imaging with intra- and peritumoral radiomic analyses and identified the 5-mm peritumoral zone as the optimal extent for enhancing the diagnostic accuracy of malignant breast nodules.16 The convergence of evidence across distinct imaging modalities further reinforces the robustness of the 5-mm margin as the preferred choice for peritumoral radiomic analysis in breast cancer.
Our second-stage findings further underscore the value of integrating multiparametric MRI in predicting ALNM. The DCE-based intra+peri5mm model achieved a training area under the curve (AUC) of 0.821, which is consistent with previously reported benchmarks. Chai et al23 reported an AUC of 0.860, and Liu et al24 described similar performance. This predictive capability is largely due to DCE-MRI’s ability to characterize tumor morphology and hemodynamic parameters. Regarding sequence optimization, most prior studies have empirically selected T2-weighted, diffusion-weighted, and contrast-enhanced sequences.25,26
Notably, Dong et al27 reported an AUC of 0.805 for a model combining T2WI and DWI. They found that T2WI superior soft tissue contrast enhances textural feature extraction, thereby improving discriminative ability. In the present study, Fs-T2WI was incorporated using only intratumoral features, as its peritumoral margin was not independently optimized. The Fs-T2WI-intra model achieved a training AUC of 0.759, indicating that T2-weighted intratumoral features encode complementary information related to morphology and edema. When integrated with DCE-derived peritumoral features in the fusion model, these T2WI intratumoral features contributed incremental predictive value, further supporting the rationale for multi-sequence integration.
Consistent with previous reports,28,29 our multivariate analysis identified maximum tumor diameter as the sole independent predictor of ALN metastasis (OR = 4.042, P < 0.001). This finding aligns with well-established oncological principles, as larger tumors are more likely to exhibit lymphatic invasion and harbor aggressive biological phenotypes. Nevertheless, the clinical model based solely on tumor diameter achieved only modest predictive performance (AUC, 0.787 in the training set; 0.743 in the internal validation set; 0.786 in the external validation set), suggesting that clinical factors alone are insufficient for accurate preoperative ALN risk stratification. This limitation motivated the incorporation of radiomic features, which capture intratumoral heterogeneity and peritumoral microenvironmental information beyond gross morphological metrics. When integrated with maximum tumor diameter, the nomogram combining the DCE-MRI-derived intra- and peritumoral (5 mm) radiomic signature yielded excellent predictive performance across all cohorts (AUCs of 0.943, 0.884, and 0.853, respectively), markedly outperforming both the clinical model and the radiomic signature alone. These results underscore the value of integrating imaging biomarkers with conventional clinical parameters, although the somewhat lower performance observed in the external validation cohort warrants careful examination of model generalizability.
The decline in model performance from the training cohort (AUC = 0.909) to the external validation cohort (AUC = 0.768) may be attributable to several factors. First, differences in image acquisition protocols, scanner hardware (although both centers used GE 3.0T systems, their hardware configurations differed), and patient populations between Center 1 and Center 2 could introduce domain shift, a well-recognized challenge when applying radiomic models to external cohorts. Second, although we adopted multiple measures to mitigate overfitting—including ICC-based feature filtering (ICC ≥ 0.75), LASSO regularization with five-fold cross-validation, and SVM hyperparameter tuning via grid search—the high dimensionality of radiomic features (hundreds of extracted features) relative to the training sample size (n = 201) may still have led to some degree of model over-optimization on the training set. Third, the relatively small external validation cohort (n = 72) may have increased the variability of performance estimates, potentially underestimating the true generalizability of the model. Collectively, these findings underscore the need for prospective multicenter studies with standardized imaging protocols to further validate the generalizability and clinical utility of the proposed model.
Strengths and Limitations
Our findings have several important clinical implications. First, the identification of the 5-mm peritumoral margin as the best-performing expansion distance among the three tested margins provides methodological guidance for future radiomic studies and supports its use for ALNM prediction. Second, the noninvasive nature of the proposed model offers a viable alternative to invasive procedures such as sentinel lymph node biopsy, potentially reducing unnecessary surgical interventions and their associated complications.
Several limitations of this study should be acknowledged. First, the external validation cohort was relatively small (n = 72). Second, peritumoral regions were defined using fixed expansion distances; adaptive or patient-specific margins merit further exploration. Third, manual segmentation is time-consuming and subject to inter-observer variability, although an ICC threshold of ≥ 0.75 was applied to ensure reproducibility. Fourth, Fs-T2WI features were extracted from the intratumoral region only, because the optimal peritumoral margin for this sequence was not independently determined; this warrants further investigation. Fifth, diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) sequences were not incorporated because of their lower spatial resolution, and their addition may further improve predictive performance. Sixth, the retrospective design may introduce selection bias, underscoring the need for prospective multicenter validation.
Conclusion
This two-stage study demonstrated that, among the three peritumoral expansion distances tested (3, 5, and 7 mm), the 5 mm margin provided the best predictive performance for axillary lymph node metastasis (ALNM) in breast cancer. The multiparametric MRI radiomic model, which integrated DCE-MRI derived intra- and peritumoral (5 mm) features with Fs-T2WI derived intratumoral features, achieved favorable predictive performance in the training, internal validation, and external validation cohorts. Moreover, the nomogram combining the radiomic signature with maximum tumor diameter exhibited excellent discrimination and clinical utility. Collectively, these findings support the potential of this noninvasive imaging-based approach for preoperative ALNM risk stratification. Nevertheless, the decrease in model performance observed in the external validation cohort underscores the need for prospective multicenter studies with larger sample sizes to confirm the generalizability of these findings before clinical implementation.
Acknowledgments
The authors thank all clinical staff for their support.
Funding Statement
This work was supported by the University Natural Science Research Project of Anhui Province (Grant No. 2024AH050711).
Ethics Approval and Informed Consent
This retrospective study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Ethics Committee of the First Hospital of Anhui University of Science and Technology (Approval No. 2025KYY039001). The requirement for informed consent was waived because the study involved only retrospective review of medical records. Patient data confidentiality was strictly maintained throughout the study.
Author Contributions
All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas. All authors took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.
Disclosure
The authors declare no conflicts of interest in this work.
References
- 1.Kim J, Harper A, McCormack V, et al. Global patterns and trends in breast cancer incidence and mortality across 185 countries. Nat Med. 2025;31(4):1154–13. doi: 10.1038/s41591-025-03502-3 [DOI] [PubMed] [Google Scholar]
- 2.Samiei S, Granzier RWY, Ibrahim A, et al. Dedicated axillary MRI-based radiomics analysis for the prediction of axillary lymph node metastasis in breast cancer. Cancers. 2021;13(4):757. doi: 10.3390/cancers13040757 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.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;318:918–926. doi: 10.1001/jama.2017.11470 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Wang L, Ding Y, Yang W, et al. A radiomics nomogram for distinguishing benign from malignant round-like breast tumors. Front Oncol. 2022;12:677803. doi: 10.3389/fonc.2022.677803 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Zhu SQ, Wang SM, Guo SQ, et al. Contrast-enhanced mammography radiomics analysis for preoperative prediction of breast cancer molecular subtypes. Acad Radiol. 2024;31(6):2228–2238. doi: 10.1016/j.acra.2023.12.005 [DOI] [PubMed] [Google Scholar]
- 6.Yao M, Ye DM, Wang Y, et al. Application of DCE-MRI radiomics and heterogeneity analysis in predicting luminal and non-luminal subtypes of breast cancer. Front Oncol. 2025;15:1523507. doi: 10.3389/fonc.2025.1523507 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Xue M, Che S, Tian Y, et al. Nomogram based on breast MRI and clinicopathologic features for predicting axillary lymph node metastasis in patients with early-stage invasive breast cancer: a retrospective study. Clin Breast Cancer. 2022;22(4):e428–e437. doi: 10.1016/j.clbc.2021.10.014 [DOI] [PubMed] [Google Scholar]
- 8.Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. 2017;14(12):749–762. doi: 10.1038/nrclinonc.2017.141 [DOI] [PubMed] [Google Scholar]
- 9.Wang S, Sun Y, Li R, et al. Diagnostic performance of perilesional radiomics analysis of contrast-enhanced mammography for the differentiation of benign and malignant breast lesions. Eur Radiol. 2022;32(1):639–649. doi: 10.1007/s00330-021-08134-y [DOI] [PubMed] [Google Scholar]
- 10.Del Corso G, Germanese D, Caudai C, et al. Adaptive machine learning approach for importance evaluation of multimodal breast cancer radiomic features. J Imaging Inform Med. 2024;37(4):1642–1651. doi: 10.1007/s10278-024-01064-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Wang Y, Shang Y, Guo Y, et al. Clinical study on the prediction of ALN metastasis based on intratumoral and peritumoral DCE-MRI radiomics and clinico-radiological characteristics in breast cancer. Front Oncol. 2024;14:1357145. doi: 10.3389/fonc.2024.1357145 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Wu PQ, Guo FL, Wang J, et al. Development and validation of a dynamic contrast-enhanced magnetic resonance imaging-based habitat and peritumoral radiomic model to predict axillary lymph node metastasis in patients with breast cancer: a retrospective study. Quant Imaging Med Surg. 2024;14(12):8211–8226. doi: 10.21037/qims-24-558 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Liu Z, Hong M, Li X, et al. Predicting axillary lymph node metastasis in breast cancer patients: a radiomics-based multicenter approach with interpretability analysis. Eur J Radiol. 2024;176:111522. doi: 10.1016/j.ejrad.2024.111522 [DOI] [PubMed] [Google Scholar]
- 14.Duan Y, Chen X, Li W, Li S, Zhang C. Multimodal radiomics and nomogram-based prediction of axillary lymph node metastasis in breast cancer: an analysis considering optimal peritumoral region. J Clin Ultrasound. 2023;51(7):1231–1241. doi: 10.1002/jcu.23520 [DOI] [PubMed] [Google Scholar]
- 15.He S, Deng B, Chen J, et al. Preoperative DBT-based radiomics for predicting axillary lymph node metastasis in breast cancer: a multi-center study. BMC Med Imaging. 2025;25(1):169. doi: 10.1186/s12880-025-01711-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Huang Z, Mo S, Wu H, et al. Optimizing breast cancer diagnosis with photoacoustic imaging: an analysis of intratumoral and peritumoral radiomics. Photoacoustics. 2024;38:100606. doi: 10.1016/j.pacs.2024.100606 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Jiang W, Meng R, Cheng Y, et al. Intra- and peritumoral based radiomics for assessment of lymphovascular invasion in invasive breast cancer. J Magn Reson Imaging. 2024;59(2):613–625. doi: 10.1002/jmri.28776 [DOI] [PubMed] [Google Scholar]
- 18.Shang Y, Wang Y, Guo Y, et al. The clinical study of intratumoral and peritumoral radiomics based on DCE-MRI for HER-2 positive and low expression prediction in breast cancer. Breast Cancer (Dove Med Press). 2024;16:957–972. doi: 10.2147/BCTT.S497770 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Zhang H, Miao Q, Fu Y, et al. Intratumoral and peritumoral radiomics based on automated breast volume scanner for predicting human epidermal growth factor receptor 2 status. Front Oncol. 2025;15:1556317. doi: 10.3389/fonc.2025.1556317 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Zhang CM, Ding ZM, Chen P, Liu QF. Value of machine learning model based on intratumoral and peritumoral DCE-MRI radiomics features in predicting axillary lymph node metastasis of breast cancer. Chin Comput Med Imaging. 2023;29(6):618–624. [Google Scholar]
- 21.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(Suppl 1):S223–S228. doi: 10.1016/j.acra.2020.10.015 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Wu J, Li Y, Gong W, et al. Multi-modality radiomics diagnosis of breast cancer based on MRI, ultrasound and mammography. BMC Med Imaging. 2025;25(1):265. doi: 10.1186/s12880-025-01767-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Chai R, Ma H, Xu M, et al. Differentiating axillary lymph node metastasis in invasive breast cancer patients: a comparison of radiomic signatures from multiparametric breast MR sequences. J Magn Reson Imaging. 2019;50(4):1125–1132. doi: 10.1002/jmri.26701 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Liu Y, Li X, Zhu L, et al. Preoperative prediction of axillary lymph node metastasis in breast cancer based on intratumoral and peritumoral DCE-MRI radiomics nomogram. Contrast Media Mol Imaging. 2022;2022:6729473. doi: 10.1155/2022/6729473 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Zhao F, Feng H, Zhu L, et al. Predictive value of multiparametric magnetic resonance imaging combined with pathological biomarkers for axillary lymph node metastasis of breast cancer. Gland Surg. 2025;14(9):1763–1776. doi: 10.21037/gs-2025-215 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Shi W, Su Y, Zhang R, et al. Prediction of axillary lymph node metastasis using a magnetic resonance imaging radiomics model of invasive breast cancer primary tumor. Cancer Imaging. 2024;24(1):122. doi: 10.1186/s40644-024-00771-y [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Dong Y, Feng Q, Yang W, et al. Preoperative prediction of sentinel lymph node metastasis in breast cancer based on radiomics of T2-weighted fat-suppression and diffusion-weighted MRI. Eur Radiol. 2018;28:582–591. doi: 10.1007/s00330-017-5005-7 [DOI] [PubMed] [Google Scholar]
- 28.Yu Y, He ZW, Ouyang J, et al. Magnetic resonance imaging radiomics predicts preoperative axillary lymph node metastasis to support surgical decisions and is associated with tumor microenvironment in invasive breast cancer: a machine learning, multicenter study. EBioMedicine. 2021;69:103460. doi: 10.1016/j.ebiom.2021.103460 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Xin Q, Yang Z. Predicting preoperative axillary lymph node metastasis to guide surgical decisions in invasive breast cancer. Front Oncol. 2026;16:1786051. doi: 10.3389/fonc.2026.1786051 [DOI] [PMC free article] [PubMed] [Google Scholar]