Comparative Analysis of 2.5D Deep Learning, 2D Deep Learning, and Radiomics Models for Predicting Axillary Lymph Node Metastasis in Breast Cancer: A Multi-Center Study

Comparative Analysis of 2.5D Deep Learning, 2D Deep Learning, and Radiomics Models for Predicting Axillary Lymph Node Metastasis in Breast Cancer: A Multi-Center Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Comparative Analysis of 2.5D Deep Learning, 2D Deep Learning, and Radiomics Models for Predicting Axillary Lymph Node Metastasis in Breast Cancer: A Multi-Center Study Lingsong Meng, Xin Zhao, Yuxia Zhang, Lin Lu, Xiang Meng, Shuangyu Li, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8372278/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 7 You are reading this latest preprint version Abstract Objective To compare the performance of 2.5D deep learning (DL) with multi-instance learning (MIL), 2D DL, and radiomics models in predicting axillary lymph node (ALN) metastasis in breast cancer (BC) patients using dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). Methods In this study, 732 patients from two independent institutions who underwent preoperative DCE-MRI were included. Based on the primary tumor region, we developed and compared four single-modality prediction models: a radiomics model, a 2D DL model, and two 2.5D DL-MIL models using different feature aggregation strategies. A stacking model was subsequently constructed by integrating the optimal single-modality models. The models’ performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and Decision Curve Analysis. Results The stacking model achieved the highest predictive performance, with AUCs of 0.962 (95% CI: 0.946–0.977) in the training set, 0.885 (95% CI: 0.837–0.933) in the internal validation set, and 0.890 (95% CI: 0.840–0.939) in the external validation set. It demonstrated high specificity (1.000 and 0.968 in the internal and external validation sets, respectively) and provided a significant net clinical benefit. The 2.5D DL-MIL models significantly outperformed the 2D DL and radiomics models. Furthermore, the stacking model maintained robust performance across key clinical subgroups defined by age, tumor size, and BI-RADS category. Conclusion The 2.5D DL-MIL-based stacking model provides an accurate, non-invasive tool for preoperative prediction of ALN metastasis in BC. Its high specificity holds promise for reducing unnecessary sentinel lymph node biopsies, thereby aiding in personalized surgical planning. Breast cancer Axillary lymph node metastasis Deep learning Magnetic resonance imaging Multi-instance learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Breast cancer (BC) is the most commonly diagnosed malignancy among women worldwide [ 1 ]. Extensive studies have confirmed that the leading cause of mortality among BC patients is tumor metastasis and recurrence [ 2 , 3 ], and the status of axillary lymph node (ALN) metastasis is strongly associated with long-term recurrence risk and mortality following local treatment [ 4 ]. Sentinel lymph node biopsy (SLNB) is the standard technique for staging ALN [ 5 ]. In cases with sentinel lymph node positivity, ALN dissection (ALND) is the conventional subsequent procedure [ 6 ]. However, SLNB is associated with procedure-related complications, including arm numbness and upper limb lymphedema [ 7 , 8 ]. Thus, there is a pressing clinical need for non-invasive, accurate preoperative prediction methods. Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), with its superior soft-tissue contrast and ability to depict lesion enhancement patterns, is valuable for evaluating ALN status [ 9 ]. Nevertheless, its diagnostic accuracy is limited by subjective radiologist interpretation and the inherent inability of visual assessment to reliably discriminate between metastatic and non-metastatic lymph nodes. Recent advances in artificial intelligence, specifically radiomics and deep learning (DL), have shown promising results in predicting lymph node metastasis from medical images. Yu et al. [ 10 ] developed an ALN-tumor radiomics signature for the preoperative prediction of ALN metastasis, which achieved an AUC of 0.870 in the external validation cohort. Li et al. [ 11 ] presented a fusion model integrating radiomics and DL features at the decision level, which achieved an AUC of 0.910 and outperformed both the traditional radiomics and DL models. Guo et al. [ 12 ] developed a convolutional recurrent neural network model based on DCE-MRI, which achieved AUCs ranging from 0.785 to 0.806 in external validation cohorts, surpassing those of traditional models. However, most of these studies rely on two-dimensional (2D) DL or hand-crafted radiomics features, which may not fully capture the spatial heterogeneity and multi-scale information inherent in tumor imaging. While three-dimensional (3D) DL approaches, by utilizing full-volume DCE-MRI data, can enhance predictive accuracy, they demand substantial computational resources and large-scale datasets for effective training [ 13 ]. To strike a balance, the 2.5D DL approach, which stacks multiple 2D slices to approximate partial 3D data, has emerged as a viable alternative: it captures more spatial context than 2D DL while remaining computationally feasible compared to 3D DL [ 14 ]. In recent years, 2.5D DL models have been successfully applied to various medical imaging tasks, including hepatocellular carcinoma recurrence prediction [ 15 ], lung cancer brain metastasis subtype differentiation [ 16 ], and clear cell renal cell carcinoma grading [ 17 ]. However, the application of this approach to BC ALN metastasis prediction remains underexplored, and its performance relative to 2D DL and radiomics has not been systematically evaluated in multi-center cohorts. To address these gaps, we incorporated multi-instance learning (MIL) into the 2.5D DL framework, a strategy that enhances the model’s ability to handle multi-slice data [ 18 ]. MIL treats each patient as a “bag” of multiple image slices (instances) and aggregates instance-level features to generate a robust patient-level prediction. This approach mitigates the risk of slice-level bias (e.g., variability in slice selection or isolated artifacts) and ensures that the model leverages the full range of multi-planar information from DCE-MRI. In this multi-center retrospective study, we aimed to systematically compare the performance of 2.5D DL-MIL, 2D DL, and radiomics models in predicting ALN metastasis in BC patients. We further developed and validated a stacking ensemble model to maximize predictive accuracy, and assessed its generalizability across independent external cohorts and its robustness across key clinical subgroups. Methods Patient population This study was approved by the Institutional Review Boards (IRBs) of our institution. Due to its retrospective nature, the requirement for written informed consent was waived. Consecutive patient data were collected from two independent clinical centers. After applying the inclusion and exclusion criteria ( Supplementary Methods ), 732 cases were enrolled. The patient selection flowchart is shown in Fig. 1 . Data from Center 1 were chronologically split into a training set (n = 411; May 2019 to May 2022) and an internal validation set (n = 176; June 2022 to May 2024). Conversely, all data from Center 2, collected during the same period as the internal validation set, served as the external validation set (n = 145). Clinical pathological data collection and reference standards Clinical and pathological data were retrieved from the electronic medical records, including: patient age, tumor histological type and grade, molecular subtype, MRI-reported ALN (MRI-ALN) status, BI-RADS category, and lesion size. Pathological lymph node results served as the reference standard. Lymph node metastasis was defined as the presence of tumor cells in the sentinel lymph node (SLN) or any ALN. Further details are provided in the Supplementary Methods . Image processing and segmentation The study workflow is depicted in Fig. 2 . All patients underwent breast MRI on a 3.0-T scanner (protocol details provided in Supplementary Methods ). The second post-contrast phase of DCE-MRI was analyzed. To mitigate resolution variability, images were resampled to an isotropic 1×1×1 mm³ voxel size and normalized using Z-score normalization. A radiologist with 15 years of breast imaging experience delineated the regions of interest (ROIs) using ITK-SNAP (v3.8.0). These ROIs were then reviewed by a second radiologist (> 20 years of experience). Both were blinded to clinical and pathological data. Given the reported value of the peri-tumoral region [ 19 ], the final ROIs were expanded outward by 10 mm for subsequent analysis. Radiomics feature extraction Radiomics features were extracted from each tumor ROI using PyRadiomics (v3.0.1). A total of 1817 features were obtained, including first-order, shape, and texture features. Extraction parameters are detailed in the Supplementary Methods . 2D DL model development and feature extraction For the 2D DL model, the largest axial cross-sectional slice of the tumor was selected as the ROI. After cropping, the slice was resized to 256×256 pixels using linear interpolation. Data augmentation, including random horizontal and vertical flips, and random cropping from 256×256 to 224×224 pixels, was applied during training to improve model generalization. The test sets were not augmented. The 2D DL model was based on a ResNet101 architecture, pre-trained on ImageNet and fine-tuned on our training set. The model was trained for 32 epochs with a learning rate of 0.01 using stochastic gradient descent (SGD) and a cross-entropy loss function. A total of 2048 deep learning features were then extracted from the penultimate average pooling layer of the trained model for subsequent analysis. 2.5D DL model development For the 2.5D DL model, the tumor’s maximal cross-sectional slice was identified on the axial, sagittal, and coronal planes. For each of these three slices, the two adjacent slices above and below were extracted, yielding a total of 9 slices per patient. This approach captures 3D contextual information and is characterized as 2.5D processing. The 2.5D DL model shared identical architectural and training configurations with the 2D DL model to ensure experimental consistency. Specifically, it was also built on the ResNet101 architecture, integrated into the same transfer learning framework, and trained using the same set of hyperparameters. Multi-instance learning and feature aggregation Multi-instance learning (MIL) was employed to derive patient-level predictions from the nine 2.5D slices. Each patient was treated as a “bag” containing nine “instances” (slices). Each slice was processed by the pre-trained ResNet101 to obtain instance-level prediction probabilities (Patch_prob) and labels (Patch_pred). Two aggregation strategies generated bag-level features: Histogram aggregation of Patch_prob and Patch_pred yielded Histo_prob and Histo_pred; and Bag-of-Words (BoW) aggregation with subsequent TF-IDF transformation of Patch_prob and Patch_pred yielded TF-IDF_prob and TF-IDF_pred. The final MIL features were concatenated as: MIL_histogram = Histo_prob ㊉ Histo_pred and MIL_TF-IDF = TF-IDF_prob ㊉ TF-IDF_pred. Details are in Supplementary Methods . Feature selection and single-modality model establishment For each feature type (radiomics, MIL_histogram, MIL_TF-IDF, 2D DL), a sequential feature selection was applied independently. Features were first normalized using Z-scores. Spearman correlation analysis removed highly correlated pairs (|r| > 0.9), retaining the feature with the higher univariate association with the outcome. For radiomics, features with an ICC > 0.80 were kept. Finally, LASSO logistic regression was used for feature selection, with the optimal λ determined by 10-fold cross-validation based on minimum binomial deviance. For single-modality model construction, each final feature subset was individually evaluated using five classifiers—LightGBM, Logistic Regression (LR), Multilayer Perceptron (MLP), Support Vector Machine (SVM) and XGBoost. The classifier yielding the highest AUC on the internal validation set for a given subset was chosen to build the corresponding predictive model. Stacking model development and model assessment The selected single-modality models were ensembled using a stacking approach to create the final integrated model. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). ROC curves, calibration curves, and decision curve analysis (DCA) were used to evaluate discrimination, calibration, and clinical utility, respectively. Interpretability of 2.5D DL model Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to improve the interpretability of our 2.5D DL framework. This approach generates localization maps that emphasize image regions critical for the classification decision. By analyzing gradient flows in the final convolutional layer of the last residual block, we visualized activation patterns corresponding to the model’s predictions, thereby identifying salient regions that contributed most significantly to the classification outcome. Statistical analysis Categorical and continuous variables were compared using the χ² test/Fisher's exact test and Mann-Whitney U test/ t -test, respectively. Model performance was evaluated by AUC, sensitivity, specificity, PPV, and NPV. A two-tailed P -value < 0.05 indicated significance. Analyses used R (v4.5.0) and Python 3.7.0 (with scikit-learn). Results Patient characteristics Table 1 summarizes the baseline characteristics of the 732 patients included in this retrospective study. The distribution of molecular subtypes and tumor grade showed no significant differences across the training, internal validation, and external validation sets (all P > 0.05). Luminal B was the predominant molecular subtype, and invasive ductal carcinoma was the most common histology. Significant differences between ALN-positive and ALN-negative patients were observed in MRI-reported ALN status and tumor size across all datasets (all P < 0.001), with positive MRI findings and larger tumor sizes being associated with ALN positivity. No significant differences were found in patient age or histology. While BI-RADS categories showed no significant differences in the training and internal validation sets, a significant disparity ( P < 0.001) was observed in the external validation set, with a higher prevalence of BI-RADS category 5 in the ALN-positive group. Table 1 Patient characteristics of three datasets Characteristic Training set (N = 411) Internal validation set (N = 176) External validation set (N = 145) ALN-negative (N = 222) ALN-positive (N = 189) P value ALN-negative (N = 90) ALN-positive (N = 86) P value ALN-negative (N = 62) ALN-positive (N = 83) P value Molecular_subtype 0.076 0.140 0.389 Luminal A 28 (13) 15 (8) 15 (17) 10 (12) 13 (21) 9 (11) Luminal B 134 (60) 137 (72) 55 (61) 6 (77) 33 (53) 51 (61) HER 2+ 31 (14) 20 (11) 10 (11) 4 (5) 8 (13) 13 (16) Triple negative 29 (13) 17 (9) 10 (11) 6 (7) 8 (13) 10 (12) Grade 0.173 0.274 0.406 1 21 (9) 9 (5) 11 4 3 (5) 1 (1) 2 163 (73) 149 (79) 68 71 38 (61) 56 (67) 3 38 (17) 31 (16) 12 11 21 (34) 26 (31) MRI_report < 0.001 < 0.001 < 0.001 No 193 (87) 89 (47) 82 (91) 33 (38) 61 (98) 36 (43) Yes 29 (13) 100 (53) 8 (9) 53 (62) 1 (2) 47 (84) Size (mm) < 0.001 < 0.001 20 129 (58) 145 (77) 46 (51) 71 (83) 23 (37) 70 (84) Histo_results 0.890 0.683 0.693 IDC 215 (97) 182 (96) 85 (94) 84 (98) 58 (94) 79 (95) ILC 6 (3) 6 (3) 3 (3) 2 (2) 3 (5) 4 (5) Others 1 (0) 1 (1) 2 (2) 0 (0) 1 (2) 0 (0) Age, (years) Median (Q1, Q3) 50 (43, 57) 50 (44, 54) 0.258 51 (43.5, 56) 51 (45, 56) 0.878 53 (44, 56) 50 (44, 56) 0.263 BI_RADS 0.211 < 0.001 < 0.001 4 104 (47) 76 (40) 50 (56) 24 (28) 42 (68) 27 (33) 5 118 (53) 113 (60) 40 (44) 62 (72) 20 (32) 56 (67) Note: ALN, axillary lymph node; IDC, invasive ductal carcinoma; ILC, invasive lobular carcinoma; BI_RADS, Breast Imaging Reporting and Data System. Feature extraction and selection A total of 1817 radiomics features were extracted, along with 100 MIL_histogram, 100 MIL_TF-IDF, and 2048 2D DL features. Following feature selection with LASSO regression, 9 radiomics, 20 MIL_histogram, 24 MIL_TF-IDF, and 23 2D DL features were retained. The relative weights of the selected features are illustrated in Supplementary Figure S1 . Analysis of model performance The MLP classifier was selected for subsequent analyses as it achieved the highest AUC on the internal validation set among all single-modality models across four feature types ( Supplementary Figure S2 ). Diagnostic performance of all models is summarized in Table 2 and Fig. 3 . In the training set, the stacking model achieved the highest AUC (0.962; 95% CI: 0.946–0.977), with no significant difference compared to the MIL_histogram model ( P = 0.103) but outperforming the MIL_TF‑IDF model ( P = 0.017). The MIL_histogram (AUC = 0.955) and MIL_TF‑IDF (AUC = 0.949) models performed similarly ( P = 0.137), both showing high specificity and PPV. In contrast, the Radiomics (AUC = 0.778) and 2D DL (AUC = 0.870) models were significantly inferior to the other three models (all P < 0.001). Table 2 Predictive Model Performance Metrics in the Datasets of This Study Model AUC (95%CI) Sensitivity Specificity PPV NPV Training set MIL_histogram 0.955 (0.938–0.972) 0.831 0.941 0.924 0.867 MIL_TF-IDF 0.949 (0.931–0.968) 0.815 0.946 0.928 0.857 Radiomics 0.778 (0.735–0.822) 0.868 0.559 0.626 0.832 2D DL 0.870 (0.837–0.903) 0.841 0.730 0.726 0.844 Stacking 0.962 (0.946–0.977) 0.884 0.905 0.888 0.901 Internal validation set MIL_histogram 0.853 (0.798–0.909) 0.686 0.944 0.922 0.759 MIL_TF-IDF 0.849 (0.792–0.905) 0.721 0.900 0.873 0.771 Radiomics 0.737 (0.663–0.811) 0.651 0.733 0.700 0.687 2D DL 0.755 (0.683–0.827) 0.791 0.656 0.687 0.766 Stacking 0.885 (0.837–0.933) 0.628 1.000 1.000 0.738 External validation set MIL_histogram 0.855 (0.794–0.915) 0.747 0.887 0.899 0.724 MIL_TF-IDF 0.871 (0.812–0.931) 0.843 0.823 0.864 0.797 Radiomics 0.772 (0.695–0.849) 0.615 0.887 0.879 0.632 2D DL 0.778 (0.700-0.856) 0.868 0.629 0.758 0.780 Stacking 0.890 (0.840–0.939) 0.639 0.968 0.964 0.667 Note: AUC, area under the receiver operating characteristic curve; CI, confidence interval; PPV, positive predictive value; NPV, negative predictive value; MIL, multi-instance learning; TF-IDF, Term Frequency Inverse Document Frequency; 2D DL, two-dimensional deep learning. In the internal validation set, the stacking model again showed the best performance (AUC = 0.885; 95% CI: 0.837–0.933), with perfect specificity and PPV (1.000), and significantly outperformed both MIL_histogram ( P = 0.013) and MIL_TF-IDF ( P = 0.009). The two MIL models remained comparable (P = 0.694), while Radiomics (AUC = 0.737) and 2D DL (AUC = 0.755) were significantly worse than the top models (all P < 0.05). In the external validation set, the stacking model achieved the highest AUC (0.890; 95% CI: 0.840–0.939), along with high specificity (0.968) and PPV (0.964). It performed significantly better than the MIL_histogram model ( P = 0.014) but not the MIL_TF-IDF model ( P = 0.241). Radiomics and 2D DL models again demonstrated lower performance. Calibration curves indicated that the stacking model was well-calibrated (Fig. 4 ), and decision curve analysis revealed a superior net clinical benefit for this model ( Supplementary Figure S3 ). Representative clinical cases are shown in Supplementary Figures S4-S5 . Subgroup-based model performance assessment Subgroup analyses were conducted to assess the robustness of the predictive models based on factors previously associated with ALN metastasis, including patient age (≤ 50 vs. >50 years), tumor size (≤ 20 vs. >20 mm), histological grade (≤ 2 vs. >2), and BI-RADS category (4 vs. 5) [ 20 , 21 ]. The results, detailed in Supplementary Figures S6-S9 , revealed that most factors (age, tumor size, and BI-RADS category) did not significantly affect the performance of any model. Notably, histological grade had a significant impact on the performance of the 2D DL model. In contrast, the radiomics, MIL_histogram, MIL_TF-IDF, and stacking models all maintained consistent and robust performance across all subgroups. Discussion This multi-center retrospective study developed and validated models based on 2.5D DL-MIL, 2D DL, and radiomics for the preoperative prediction of ALN metastasis in BC patients using the second phase of DCE-MRI. A stacking ensemble model was then constructed using the optimal single-modality models. The results demonstrated that the stacking model achieved the highest diagnostic performance, with AUCs of 0.962 (95% CI: 0.946–0.977), 0.885 (95% CI: 0.837–0.933), and 0.890 (95% CI: 0.840–0.939) in the training, internal validation, and external validation sets, respectively. It also exhibited robust generalizability, satisfactory calibration, and a substantial net clinical benefit. Subgroup analyses confirmed the model's stable performance across patient subgroups stratified by age, tumor size, nuclear grade, and BI-RADS category. Preoperative prediction of occult axillary lymph node (ALN) metastasis in breast cancer remains a significant clinical challenge. The Memorial Sloan Kettering Cancer Center (MSKCC) nomogram, the most widely used predictive tool, integrates clinicopathological parameters such as tumor size, lymphovascular invasion, and sentinel lymph node status to estimate non-sentinel lymph node metastasis risk [ 22 ]. However, it lacks integration of quantitative imaging predictors. Although MRI features of primary breast cancer have demonstrated association with ALN metastasis and offer potential for improved preoperative prediction [ 23 – 25 ], their qualitative assessment is operator-dependent and fails to comprehensively capture imaging information. Recent radiomics approaches have attempted to address this limitation [ 4 , 26 ], but feature inconsistency across models hinders reproducibility. This generalizability concern is evidenced in our external validation, where the radiomics model achieved an AUC of only 0.772 (95% CI: 0.695–0.849). Deep learning (DL) approaches for lymph node assessment have been applied in several solid tumors. Multiple studies have investigated DL models for predicting ALN metastasis in BC patients [ 11 , 12 , 21 , 27 ]. For example, Chen et al. [ 21 ] developed a DCE-MRI-based CNN model to predict SLN and non-SLN metastasis in breast cancer, reporting AUCs of 0.899 (internal validation), 0.885 (external test 1), and 0.768 (external test 2) for SLN prediction, and 0.800, 0.763, and 0.728 for non-SLN prediction in the same cohorts. Guo et al. [ 27 ] developed multimodal models combining radiomics and DL features from mammography and MRI, with their combined MLP model achieving the highest test AUC of 0.846. However, most existing studies rely on 2D or 3D DL architectures. In our study, we proposed two 2.5D DL-MIL models based on DCE-MRI images of the primary tumor. Both models achieved promising and consistent performance, with AUCs ranging from 0.849 to 0.871 across the validation sets. Our findings confirm the superiority of the 2.5D DL-MIL approach over the 2D DL model, aligning with previous studies [ 28 ]. This improvement stems from several factors. Unlike 2D models relying on a single cross-sectional slice, the 2.5D DL-MIL model incorporates partial 3D context by aggregating multi-slice features across three anatomical planes using MIL. This multi-planar sampling preserves key spatial relationships between the tumor and its microenvironment—features associated with ALN metastasis, such as peritumoral angiogenesis. Moreover, MIL reduces inter-slice variability and selection bias by treating each patient as a “bag” of instances, yielding a robust patient-level representation less affected by slice-specific noise. In contrast, the radiomics model showed the lowest performance (external validation AUC = 0.772). Handcrafted radiomics features, based on predefined mathematical descriptors, often fail to capture complex, nonlinear tumor heterogeneity patterns linked to ALN metastasis. Deep learning models, however, automatically learn discriminative high-level features from imaging data, capturing subtle patterns beyond conventional radiomics and human perception. The high specificity of the stacking model (1.000 internally, 0.968 externally) holds significant clinical promise. Given that SLNB is associated with complications like arm numbness and lymphedema [ 7 , 8 ], a model that minimizes false positives is crucial. In our external validation, 96.8% of patients predicted as ALN-negative were pathologically negative, suggesting that only 3.2% might undergo unnecessary invasive staging. This aligns with the de-escalation goals of trials like ACOSOG Z0011 [ 29 ]. Additionally, the model's consistent performance across subgroups, including challenging cases like small tumors (≤ 20 mm) where it maintained an AUC > 0.870, addresses a key limitation of many existing prediction tools. Several limitations of this study warrant consideration. First, its retrospective nature may introduce selection bias; prospective validation in a consecutive cohort is needed to confirm real-world clinical utility. Second, the model was trained exclusively on the second post-contrast DCE-MRI phase; integrating additional sequences (e.g., DWI) might improve performance. Third, the external validation was performed with data from a single additional center; future multi-center studies with larger, more diverse cohorts (e.g., including 1.5-T MRI data) are necessary to enhance generalizability. Finally, although ROIs were delineated by experienced radiologists and reviewed to minimize variability, inter-observer differences in segmentation could influence feature extraction. Future work could explore automated segmentation to standardize this process. Conclusion The 2.5D DL-MIL-based stacking model provides a non-invasive, accurate, and robust tool for the preoperative prediction of ALN metastasis in BC. By integrating complementary features from multiple approaches, it outperforms single-modality models and demonstrates consistent efficacy across diverse patient subgroups. Its high specificity has the potential to reduce unnecessary invasive procedures, supporting its broader clinical adoption. Future prospective, multi-center studies integrating multi-parametric MRI data are warranted to further validate and refine this tool for personalized treatment planning. Abbreviations ALN Axillary lymph node ALND Axillary lymph node dissection SLNB Sentinel lymph node biopsy DCE-MRI Dynamic contrast-enhanced magnetic resonance imaging ROI Region of interest 2D/3D Two-dimensional/three-dimensional DL Deep learning MIL Multi-instance learning BI-RADS Breast Imaging Reporting and Data System AUC Area under the curve ICC Intraclass correlation coefficient LASSO Least absolute shrinkage and selection operator ROC Receiver operating characteristic DCA Decision curve analysis. Declarations Ethics declaration This retrospective study was conducted in accordance with the Declaration of Helsinki. The study was approved by the Ethics Committee of Shangqiu Medical College (approval number: SYLL2025-008). Given the use of anonymized data, informed consent was waived for this study. Clinical trial number: Not applicable. Consent for publication As all patient data utilized in this retrospective study were fully anonymized and could not be traced back to individual participants, informed consent for publication was formally waived. Competing interests The authors declare no competing interests. Funding This work was supported by the Science and Technology Research Project of Henan Province (252102310061), Key Scientific Research Projects of Higher Education Institutions of Henan Province (25B320031, 26B320017), Science and Technology Program of Shangqiu City (2024075, 2025068), and Doctoral Scientific Research Initiation Project of Shangqiu Medical College (BSJH004). Additional support was provided by the Henan Provincial Health Commission Key Laboratory of Imaging Precision Diagnosis. Author Contribution LSM: Data curation, Writing – Original draft preparation. XZ: Conceptualization, Supervision, Visualization. YXZ: Conceptualization, Methodology, Supervision. LL: Conceptualization, Supervision, Visualization. XM: Conceptualization, Writing – Review & Editing. SYL: Resources, Formal analysis. FMS: Visualization, Formal analysis. XAZ: Conceptualization, Visualization, Writing – Review & Editing. All authors have read and approved the final manuscript. Data Availability The datasets employed and/or examined in this study can be obtained from the corresponding author upon reasonable request. References Siegel RL, Miller KD, Wagle NS, Jemal A. Cancer statistics, 2023. CA: A Cancer. J Clin. 2023;73:17–48. DeSantis CE, Ma J, Gaudet MM, Newman LA, Miller KD, Goding Sauer A, et al. Breast cancer statistics, 2019. CA: A Cancer. J Clin. 2019;69:438–51. Siegel RL, Miller KD, Jemal A, Cancer statistics. 2017. CA: A Cancer Journal for Clinicians. 2017;67:7–30. Yu Y, Tan Y, Xie C, Hu Q, Ouyang J, Chen Y, et al. Development and Validation of a Preoperative Magnetic Resonance Imaging Radiomics-Based Signature to Predict Axillary Lymph Node Metastasis and Disease-Free Survival in Patients With Early-Stage Breast Cancer. JAMA Netw open. 2020;3:e2028086. Gentilini O, Veronesi U. Abandoning sentinel lymph node biopsy in early breast cancer? A new trial in progress at the European Institute of Oncology of Milan (SOUND: Sentinel node vs Observation after axillary UltraSouND). Breast (Edinburgh, Scotland). 2012;21:678–681. Liang Y, Chen X, Tong Y, Zhan W, Zhu Y, Wu J, et al. Higher axillary lymph node metastasis burden in breast cancer patients with positive preoperative node biopsy: may not be appropriate to receive sentinel lymph node biopsy in the post-ACOSOG Z0011 trial era. World J Surg Oncol. 2019;17:37. Boughey JC, Moriarty JP, Degnim AC, Gregg MS, Egginton JS, Long KH. Cost modeling of preoperative axillary ultrasound and fine-needle aspiration to guide surgery for invasive breast cancer. Ann Surg Oncol. 2010;17:953–8. Langer I, Guller U, Berclaz G, Koechli OR, Schaer G, Fehr MK, 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. Le Boulc'h M, Gilhodes J, Steinmeyer Z, Molière S, Mathelin C. Pretherapeutic Imaging for Axillary Staging in Breast Cancer: A Systematic Review and Meta-Analysis of Ultrasound, MRI and FDG PET. J Clin Med. 2021;10(7):1543. Yu Y, He Z, Ouyang J, Tan Y, Chen Y, Gu Y, 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. Li X, Yang L, Jiao X. Comparison of Traditional Radiomics, Deep Learning Radiomics and Fusion Methods for Axillary Lymph Node Metastasis Prediction in Breast Cancer. Acad Radiol. 2023;30:1281–7. Guo YJ, Yin R, Zhang Q, Han JQ, Dou ZX, Wang PB, et al. MRI-Based Kinetic Heterogeneity Evaluation in the Accurate Access of Axillary Lymph Node Status in Breast Cancer Using a Hybrid CNN‐RNN Model. J Magn Reson Imaging. 2024;60:1352–64. Zhang Y, Liao Q, Ding L, Zhang J. Bridging 2D and 3D segmentation networks for computation-efficient volumetric medical image segmentation: An empirical study of 2.5D solutions. Comput Med Imaging Graph. 2022;99:102088. Zeng Y, Zhang X, Kawasumi Y, Usui A, Ichiji K, Funayama M, et al. A 2.5D Deep Learning-Based Method for Drowning Diagnosis Using Post-Mortem Computed Tomography. IEEE J biomedical health Inf. 2023;27:1026–35. Zhang Y-B, Chen Z-Q, Bu Y, Lei P, Yang W, Zhang W. Construction of a 2.5D Deep Learning Model for Predicting Early Postoperative Recurrence of Hepatocellular Carcinoma Using Multi-View and Multi-Phase CT Images. J Hepatocellular Carcinoma. 2024;11:2223–39. Zhu J, Zou L, Xie X, Xu R, Tian Y, Zhang B. 2.5D deep learning based on multi-parameter MRI to differentiate primary lung cancer pathological subtypes in patients with brain metastases. Eur J Radiol. 2024;180:111712. Yang Z, Jiang H, Shan S, Wang X, Kou Q, Wang C, et al. 2.5D Deep Learning-Based Prediction of Pathological Grading of Clear Cell Renal Cell Carcinoma Using Contrast-Enhanced CT: A Multicenter Study. Acad Radiol. 2025;32:5907–16. He J, Xu J, Chen W, Cao M, Zhang J, Yang Q, et al. Development of a deep learning model for T1N0 gastric cancer diagnosis using 2.5D radiomic data in preoperative CT images. npj Precision Oncol. 2025;9(1):249. Liu Z, Hong M, Li X, Lin L, Tan X, Liu Y. Predicting axillary lymph node metastasis in breast cancer patients: A radiomics-based multicenter approach with interpretability analysis. Eur J Radiol. 2024;176:111522. Liu H, Zou L, Xu N, Shen H, Zhang Y, Wan P, et al. Deep learning radiomics based prediction of axillary lymph node metastasis in breast cancer. npj Breast Cancer. 2024;10(1):22. Chen M, Kong C, Lin G, Chen W, Guo X, Chen Y, et al. Development and validation of convolutional neural network-based model to predict the risk of sentinel or non-sentinel lymph node metastasis in patients with breast cancer: a machine learning study. eClinicalMedicine. 2023;63:102176. Memorial Sloan Kettering Cancer Center breast cancer nomogram. http://nomo grams.mskcc.org/breast/index.aspx . Accessed June 18, 2019. Luo H, Chen Z, Xu H, Ren J, Zhou P. Peritumoral edema enhances MRI-based deep learning radiomic model for axillary lymph node metastasis burden prediction in breast cancer. Sci Rep. 2024;14:18900. Baltzer PA, Dietzel M, Burmeister HP, Zoubi R, Gajda M, Camara O, et al. Application of MR mammography beyond local staging: is there a potential to accurately assess axillary lymph nodes? evaluation of an extended protocol in an initial prospective study. AJR Am J Roentgenol. 2011;196:W641–7. Iima M, Honda M, Sigmund EE, Ohno Kishimoto A, Kataoka M, Togashi K. Diffusion MRI of the breast: Current status and future directions. J Magn Reson Imaging. 2020;52:70–90. Han L, Zhu Y, Liu Z, Yu T, He C, Jiang W, et al. Radiomic nomogram for prediction of axillary lymph node metastasis in breast cancer. Eur Radiol. 2019;29:3820–9. Guo F, Sun S, Deng X, Wang Y, Yao W, Yue P, et al. Predicting axillary lymph node metastasis in breast cancer using a multimodal radiomics and deep learning model. Front Immunol. 2024;15:1482020. Mai W, Fan X, Zhang L, Li J, Chen L, Hua X, et al. Deep learning for differential diagnosis of parotid tumors based on 2.5D magnetic resonance imaging. Ann Med. 2025;57(1):2520401. Giuliano AE, Ballman KV, McCall L, Beitsch PD, Brennan MB, Kelemen PR, 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. Description. of each illustration. Additional Declarations No competing interests reported. Supplementary Files Supplementarymaterials.doc Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 21 Jan, 2026 Reviewers agreed at journal 14 Jan, 2026 Reviewers invited by journal 14 Jan, 2026 Editor assigned by journal 12 Jan, 2026 Editor invited by journal 18 Dec, 2025 Submission checks completed at journal 18 Dec, 2025 First submitted to journal 18 Dec, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8372278","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":574691517,"identity":"97be8a70-9a63-46de-a685-397a79bc273b","order_by":0,"name":"Lingsong Meng","email":"","orcid":"","institution":"Shangqiu Medical College","correspondingAuthor":false,"prefix":"","firstName":"Lingsong","middleName":"","lastName":"Meng","suffix":""},{"id":574691518,"identity":"2a89b852-6fd4-4f5f-9945-7030e4f4d8d7","order_by":1,"name":"Xin Zhao","email":"","orcid":"","institution":"The Third Affiliated Hospital of Zhengzhou University","correspondingAuthor":false,"prefix":"","firstName":"Xin","middleName":"","lastName":"Zhao","suffix":""},{"id":574691519,"identity":"7010a4d9-c906-4dd9-87d3-0fd60a160819","order_by":2,"name":"Yuxia Zhang","email":"","orcid":"","institution":"Shangqiu Medical College","correspondingAuthor":false,"prefix":"","firstName":"Yuxia","middleName":"","lastName":"Zhang","suffix":""},{"id":574691520,"identity":"8fd3f1fc-9ab7-41df-a9c0-79e5024e30a8","order_by":3,"name":"Lin Lu","email":"","orcid":"","institution":"The Third Affiliated Hospital of Zhengzhou University","correspondingAuthor":false,"prefix":"","firstName":"Lin","middleName":"","lastName":"Lu","suffix":""},{"id":574691521,"identity":"0fc992a6-9888-4d92-b342-78dfe02f4db0","order_by":4,"name":"Xiang Meng","email":"","orcid":"","institution":"Shangqiu Medical College","correspondingAuthor":false,"prefix":"","firstName":"Xiang","middleName":"","lastName":"Meng","suffix":""},{"id":574691522,"identity":"507bd8ad-3812-4cfd-b846-f7657d1058b4","order_by":5,"name":"Shuangyu Li","email":"","orcid":"","institution":"The First People’s Hospital of Shangqiu City","correspondingAuthor":false,"prefix":"","firstName":"Shuangyu","middleName":"","lastName":"Li","suffix":""},{"id":574691523,"identity":"9e0d0985-0f52-45db-8953-8af3895ead8c","order_by":6,"name":"Fuming Shao","email":"","orcid":"","institution":"The Second Affiliated Hospital of Shangqiu Medical College","correspondingAuthor":false,"prefix":"","firstName":"Fuming","middleName":"","lastName":"Shao","suffix":""},{"id":574691524,"identity":"498ae39a-d330-43f0-ac70-07ad1940002a","order_by":7,"name":"Xiaoan Zhang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAp0lEQVRIiWNgGAWjYHACxgcQOoF4LcwGJGthkyBNi/yM5GeVP/4cZuBnzzFg+LmDCC2MPcfMbkjwHGaQ7HljwNh7hggtzOwNZjcMJA4zGNzIMWBmbCNCCxsz+7eCBIPDDPZEa+Fh7zFjOJAAtEWCWC0SPGeKJRsOpPNInHlWcLCXGC3yM9I3fvzxx1qOvz1544OfxGiBOxBEHCBBwygYBaNgFIwCfAAAq9YvZQEKtz0AAAAASUVORK5CYII=","orcid":"","institution":"The Third Affiliated Hospital of Zhengzhou University","correspondingAuthor":true,"prefix":"","firstName":"Xiaoan","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2025-12-16 05:54:09","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8372278/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8372278/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":100600228,"identity":"dec6d495-0ec5-41e9-af12-7bbdab02c850","added_by":"auto","created_at":"2026-01-19 14:46:58","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":40562,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/cf2d38547593edbf18b5e24b.docx"},{"id":100600185,"identity":"bffd6224-fb60-43dd-96b1-085b0af84489","added_by":"auto","created_at":"2026-01-19 14:46:54","extension":"tif","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":912976,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.tif","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/71408052fbf7b048478e0076.tif"},{"id":100600408,"identity":"40faf5b1-a85f-4327-bd4f-857501884fa3","added_by":"auto","created_at":"2026-01-19 14:47:58","extension":"doc","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":84480,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.doc","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/0c5cd83ba16a3f7050383a27.doc"},{"id":100600458,"identity":"6402e76f-03f0-43d1-8f86-c6377dc54fae","added_by":"auto","created_at":"2026-01-19 14:48:32","extension":"tif","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7816060,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.tif","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/d8c68cc3d5e1d6d9a5a4f97e.tif"},{"id":100600629,"identity":"6d30638b-c95b-489c-89bd-f7f552f067f7","added_by":"auto","created_at":"2026-01-19 14:49:10","extension":"doc","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":41984,"visible":true,"origin":"","legend":"","description":"","filename":"Table2.doc","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/c161d6f6c4ab71e50b84b3bb.doc"},{"id":100600160,"identity":"9f32e398-5e23-45bd-b2f7-4e97312612e4","added_by":"auto","created_at":"2026-01-19 14:46:43","extension":"tif","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":870152,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.tif","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/ff3ba23416c8b365c9791e33.tif"},{"id":100600319,"identity":"2b79563b-7bf4-44b0-827e-4a40dbe1653e","added_by":"auto","created_at":"2026-01-19 14:47:25","extension":"tif","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":360136,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.tif","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/5bc98de6117d686b1b8491c4.tif"},{"id":100600110,"identity":"911e63a6-ff14-40c5-a658-27c079fe9fa7","added_by":"auto","created_at":"2026-01-19 14:46:20","extension":"json","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9276,"visible":true,"origin":"","legend":"","description":"","filename":"d5dbd6fa172f460eb51bb7fd3b5a7a09.json","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/908f72b884b5b95b24d2ef68.json"},{"id":100600229,"identity":"c5998871-2860-4972-9776-0aae29ab95eb","added_by":"auto","created_at":"2026-01-19 14:46:59","extension":"doc","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":18488320,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterials.doc","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/04bb21ccb4915a28152b4232.doc"},{"id":100600393,"identity":"97a74a0c-793e-4cfe-8bdf-90afc3e9b3e3","added_by":"auto","created_at":"2026-01-19 14:47:48","extension":"xml","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":115651,"visible":true,"origin":"","legend":"","description":"","filename":"d5dbd6fa172f460eb51bb7fd3b5a7a091enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/bf9bdce282bc72de0a41d963.xml"},{"id":100600240,"identity":"bda40eeb-9672-4060-836c-4ff356928164","added_by":"auto","created_at":"2026-01-19 14:47:04","extension":"tif","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":912976,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.tif","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/dc7c35c081fb24c53d81f838.tif"},{"id":100600440,"identity":"b51d4a15-eeee-4c44-bd33-c6569fc07e13","added_by":"auto","created_at":"2026-01-19 14:48:16","extension":"tif","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7816060,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.tif","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/c5a8b49fa49d9b2d930f7ac1.tif"},{"id":100600131,"identity":"bdbc7799-4550-4fc6-a5f4-1ba574dc41dd","added_by":"auto","created_at":"2026-01-19 14:46:30","extension":"tif","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":870152,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.tif","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/a28cee1480caf0fac5335100.tif"},{"id":100600114,"identity":"f500bd19-fc3c-4df8-9dc5-3d726218dc30","added_by":"auto","created_at":"2026-01-19 14:46:21","extension":"tif","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":360136,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.tif","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/f0a8db2ad3fdea2da5408df2.tif"},{"id":100600117,"identity":"08a5f546-9f44-4734-aa90-90da0c986915","added_by":"auto","created_at":"2026-01-19 14:46:22","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":388383,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/b20ce7fd0b6b38e9330f7dec.png"},{"id":100600138,"identity":"9997ea8f-65cd-44d8-b282-2a93fb888609","added_by":"auto","created_at":"2026-01-19 14:46:36","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":851732,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/4db648163f2e73391483d08a.png"},{"id":100600387,"identity":"74b35c8d-2748-4f63-8016-d5087566d5e9","added_by":"auto","created_at":"2026-01-19 14:47:45","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":131511,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/3d678b27c36bdc0b00b216d6.png"},{"id":100600344,"identity":"01dfa766-e500-4a37-94e0-59bd5d37c952","added_by":"auto","created_at":"2026-01-19 14:47:39","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":65186,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/5e5a57fb78710baa466825f8.png"},{"id":100600258,"identity":"583a9c3f-37d2-4739-8097-8ba2cc7fd9d5","added_by":"auto","created_at":"2026-01-19 14:47:15","extension":"xml","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":114629,"visible":true,"origin":"","legend":"","description":"","filename":"d5dbd6fa172f460eb51bb7fd3b5a7a091structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/2f0b44c76439f966cf8374c0.xml"},{"id":100600384,"identity":"94f18011-accd-41db-85ee-0928821eb07f","added_by":"auto","created_at":"2026-01-19 14:47:43","extension":"html","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":125658,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/201e3e5885e2b6852086571a.html"},{"id":100600130,"identity":"3f95c829-9f54-4a7b-a9ca-8cbd171674e5","added_by":"auto","created_at":"2026-01-19 14:46:30","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1555196,"visible":true,"origin":"","legend":"\u003cp\u003eThe patient inclusion flowchart.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/9104a77d8d5e688c3dea1239.png"},{"id":100599917,"identity":"fce8bec3-d580-4e82-9e3a-82ec2bb15e3e","added_by":"auto","created_at":"2026-01-19 14:45:36","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":2202038,"visible":true,"origin":"","legend":"\u003cp\u003eThe study’s flowchart.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/830c5e54e93e4af0efa7dafe.png"},{"id":100600226,"identity":"aba014ec-9ef3-49af-9370-0190e36e15a9","added_by":"auto","created_at":"2026-01-19 14:46:58","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":2585672,"visible":true,"origin":"","legend":"\u003cp\u003eThe ROC curves and Delong test in this study. Panels A-C present the receiver operating characteristic (ROC) curves for the training, internal validation, and external validation sets, respectively. The corresponding Delong tests for statistically comparing the models across these three datasets are displayed in Panels D-F.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/930ac41b9437c7a81363ee9a.png"},{"id":100600210,"identity":"c5f032aa-7a3a-4162-8d3a-5bfced06d4d5","added_by":"auto","created_at":"2026-01-19 14:46:55","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1127005,"visible":true,"origin":"","legend":"\u003cp\u003eThe calibration curves in this study. Panels A-C present the calibration curves for the training, internal validation, and external validation sets, respectively.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/75e858688aecdcb59b2a555b.png"},{"id":100803978,"identity":"18a6fee4-b02c-44aa-880d-3e3c373cc522","added_by":"auto","created_at":"2026-01-21 14:33:19","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":15890190,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/28401c8b-9b0b-4dca-b393-313cb6113665.pdf"},{"id":100599938,"identity":"daec21b3-851b-4f27-9341-d1776e3bcbee","added_by":"auto","created_at":"2026-01-19 14:45:44","extension":"doc","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":18488320,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterials.doc","url":"https://assets-eu.researchsquare.com/files/rs-8372278/v1/e8a6e4fe6256349b488a9ad6.doc"}],"financialInterests":"No competing interests reported.","formattedTitle":"Comparative Analysis of 2.5D Deep Learning, 2D Deep Learning, and Radiomics Models for Predicting Axillary Lymph Node Metastasis in Breast Cancer: A Multi-Center Study","fulltext":[{"header":"Introduction","content":"\u003cp\u003eBreast cancer (BC) is the most commonly diagnosed malignancy among women worldwide [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Extensive studies have confirmed that the leading cause of mortality among BC patients is tumor metastasis and recurrence [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], and the status of axillary lymph node (ALN) metastasis is strongly associated with long-term recurrence risk and mortality following local treatment [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Sentinel lymph node biopsy (SLNB) is the standard technique for staging ALN [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. In cases with sentinel lymph node positivity, ALN dissection (ALND) is the conventional subsequent procedure [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. However, SLNB is associated with procedure-related complications, including arm numbness and upper limb lymphedema [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Thus, there is a pressing clinical need for non-invasive, accurate preoperative prediction methods.\u003c/p\u003e \u003cp\u003eDynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), with its superior soft-tissue contrast and ability to depict lesion enhancement patterns, is valuable for evaluating ALN status [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Nevertheless, its diagnostic accuracy is limited by subjective radiologist interpretation and the inherent inability of visual assessment to reliably discriminate between metastatic and non-metastatic lymph nodes.\u003c/p\u003e \u003cp\u003eRecent advances in artificial intelligence, specifically radiomics and deep learning (DL), have shown promising results in predicting lymph node metastasis from medical images. Yu et al. [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] developed an ALN-tumor radiomics signature for the preoperative prediction of ALN metastasis, which achieved an AUC of 0.870 in the external validation cohort. Li et al. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] presented a fusion model integrating radiomics and DL features at the decision level, which achieved an AUC of 0.910 and outperformed both the traditional radiomics and DL models. Guo et al. [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] developed a convolutional recurrent neural network model based on DCE-MRI, which achieved AUCs ranging from 0.785 to 0.806 in external validation cohorts, surpassing those of traditional models. However, most of these studies rely on two-dimensional (2D) DL or hand-crafted radiomics features, which may not fully capture the spatial heterogeneity and multi-scale information inherent in tumor imaging. While three-dimensional (3D) DL approaches, by utilizing full-volume DCE-MRI data, can enhance predictive accuracy, they demand substantial computational resources and large-scale datasets for effective training [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. To strike a balance, the 2.5D DL approach, which stacks multiple 2D slices to approximate partial 3D data, has emerged as a viable alternative: it captures more spatial context than 2D DL while remaining computationally feasible compared to 3D DL [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. In recent years, 2.5D DL models have been successfully applied to various medical imaging tasks, including hepatocellular carcinoma recurrence prediction [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], lung cancer brain metastasis subtype differentiation [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], and clear cell renal cell carcinoma grading [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. However, the application of this approach to BC ALN metastasis prediction remains underexplored, and its performance relative to 2D DL and radiomics has not been systematically evaluated in multi-center cohorts.\u003c/p\u003e \u003cp\u003eTo address these gaps, we incorporated multi-instance learning (MIL) into the 2.5D DL framework, a strategy that enhances the model\u0026rsquo;s ability to handle multi-slice data [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. MIL treats each patient as a \u0026ldquo;bag\u0026rdquo; of multiple image slices (instances) and aggregates instance-level features to generate a robust patient-level prediction. This approach mitigates the risk of slice-level bias (e.g., variability in slice selection or isolated artifacts) and ensures that the model leverages the full range of multi-planar information from DCE-MRI.\u003c/p\u003e \u003cp\u003eIn this multi-center retrospective study, we aimed to systematically compare the performance of 2.5D DL-MIL, 2D DL, and radiomics models in predicting ALN metastasis in BC patients. We further developed and validated a stacking ensemble model to maximize predictive accuracy, and assessed its generalizability across independent external cohorts and its robustness across key clinical subgroups.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003ePatient population\u003c/h2\u003e \u003cp\u003eThis study was approved by the Institutional Review Boards (IRBs) of our institution. Due to its retrospective nature, the requirement for written informed consent was waived. Consecutive patient data were collected from two independent clinical centers. After applying the inclusion and exclusion criteria (\u003cb\u003eSupplementary Methods\u003c/b\u003e), 732 cases were enrolled. The patient selection flowchart is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Data from Center 1 were chronologically split into a training set (n\u0026thinsp;=\u0026thinsp;411; May 2019 to May 2022) and an internal validation set (n\u0026thinsp;=\u0026thinsp;176; June 2022 to May 2024). Conversely, all data from Center 2, collected during the same period as the internal validation set, served as the external validation set (n\u0026thinsp;=\u0026thinsp;145).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eClinical pathological data collection and reference standards\u003c/h3\u003e\n\u003cp\u003eClinical and pathological data were retrieved from the electronic medical records, including: patient age, tumor histological type and grade, molecular subtype, MRI-reported ALN (MRI-ALN) status, BI-RADS category, and lesion size. Pathological lymph node results served as the reference standard. Lymph node metastasis was defined as the presence of tumor cells in the sentinel lymph node (SLN) or any ALN. Further details are provided in the \u003cb\u003eSupplementary Methods\u003c/b\u003e.\u003c/p\u003e\n\u003ch3\u003eImage processing and segmentation\u003c/h3\u003e\n\u003cp\u003eThe study workflow is depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. All patients underwent breast MRI on a 3.0-T scanner (protocol details provided in \u003cb\u003eSupplementary Methods\u003c/b\u003e). The second post-contrast phase of DCE-MRI was analyzed. To mitigate resolution variability, images were resampled to an isotropic 1\u0026times;1\u0026times;1 mm\u0026sup3; voxel size and normalized using Z-score normalization.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eA radiologist with 15 years of breast imaging experience delineated the regions of interest (ROIs) using ITK-SNAP (v3.8.0). These ROIs were then reviewed by a second radiologist (\u0026gt;\u0026thinsp;20 years of experience). Both were blinded to clinical and pathological data. Given the reported value of the peri-tumoral region [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], the final ROIs were expanded outward by 10 mm for subsequent analysis.\u003c/p\u003e\n\u003ch3\u003eRadiomics feature extraction\u003c/h3\u003e\n\u003cp\u003eRadiomics features were extracted from each tumor ROI using PyRadiomics (v3.0.1). A total of 1817 features were obtained, including first-order, shape, and texture features. Extraction parameters are detailed in the \u003cb\u003eSupplementary Methods\u003c/b\u003e.\u003c/p\u003e \u003cp\u003e \u003cb\u003e2D DL model development and feature extraction\u003c/b\u003e \u003c/p\u003e \u003cp\u003eFor the 2D DL model, the largest axial cross-sectional slice of the tumor was selected as the ROI. After cropping, the slice was resized to 256\u0026times;256 pixels using linear interpolation. Data augmentation, including random horizontal and vertical flips, and random cropping from 256\u0026times;256 to 224\u0026times;224 pixels, was applied during training to improve model generalization. The test sets were not augmented.\u003c/p\u003e \u003cp\u003eThe 2D DL model was based on a ResNet101 architecture, pre-trained on ImageNet and fine-tuned on our training set. The model was trained for 32 epochs with a learning rate of 0.01 using stochastic gradient descent (SGD) and a cross-entropy loss function. A total of 2048 deep learning features were then extracted from the penultimate average pooling layer of the trained model for subsequent analysis.\u003c/p\u003e \u003cp\u003e \u003cb\u003e2.5D DL model development\u003c/b\u003e \u003c/p\u003e \u003cp\u003eFor the 2.5D DL model, the tumor\u0026rsquo;s maximal cross-sectional slice was identified on the axial, sagittal, and coronal planes. For each of these three slices, the two adjacent slices above and below were extracted, yielding a total of 9 slices per patient. This approach captures 3D contextual information and is characterized as 2.5D processing.\u003c/p\u003e \u003cp\u003eThe 2.5D DL model shared identical architectural and training configurations with the 2D DL model to ensure experimental consistency. Specifically, it was also built on the ResNet101 architecture, integrated into the same transfer learning framework, and trained using the same set of hyperparameters.\u003c/p\u003e\n\u003ch3\u003eMulti-instance learning and feature aggregation\u003c/h3\u003e\n\u003cp\u003eMulti-instance learning (MIL) was employed to derive patient-level predictions from the nine 2.5D slices. Each patient was treated as a \u0026ldquo;bag\u0026rdquo; containing nine \u0026ldquo;instances\u0026rdquo; (slices). Each slice was processed by the pre-trained ResNet101 to obtain instance-level prediction probabilities (Patch_prob) and labels (Patch_pred).\u003c/p\u003e \u003cp\u003eTwo aggregation strategies generated bag-level features: Histogram aggregation of Patch_prob and Patch_pred yielded Histo_prob and Histo_pred; and Bag-of-Words (BoW) aggregation with subsequent TF-IDF transformation of Patch_prob and Patch_pred yielded TF-IDF_prob and TF-IDF_pred. The final MIL features were concatenated as: MIL_histogram\u0026thinsp;=\u0026thinsp;Histo_prob ㊉ Histo_pred and MIL_TF-IDF\u0026thinsp;=\u0026thinsp;TF-IDF_prob ㊉ TF-IDF_pred. Details are in \u003cb\u003eSupplementary Methods\u003c/b\u003e.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eFeature selection and single-modality model establishment\u003c/h2\u003e \u003cp\u003eFor each feature type (radiomics, MIL_histogram, MIL_TF-IDF, 2D DL), a sequential feature selection was applied independently. Features were first normalized using Z-scores. Spearman correlation analysis removed highly correlated pairs (|r| \u0026gt; 0.9), retaining the feature with the higher univariate association with the outcome. For radiomics, features with an ICC\u0026thinsp;\u0026gt;\u0026thinsp;0.80 were kept. Finally, LASSO logistic regression was used for feature selection, with the optimal λ determined by 10-fold cross-validation based on minimum binomial deviance.\u003c/p\u003e \u003cp\u003eFor single-modality model construction, each final feature subset was individually evaluated using five classifiers\u0026mdash;LightGBM, Logistic Regression (LR), Multilayer Perceptron (MLP), Support Vector Machine (SVM) and XGBoost. The classifier yielding the highest AUC on the internal validation set for a given subset was chosen to build the corresponding predictive model.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eStacking model development and model assessment\u003c/h3\u003e\n\u003cp\u003eThe selected single-modality models were ensembled using a stacking approach to create the final integrated model. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). ROC curves, calibration curves, and decision curve analysis (DCA) were used to evaluate discrimination, calibration, and clinical utility, respectively.\u003c/p\u003e\n\u003ch3\u003eInterpretability of 2.5D DL model\u003c/h3\u003e\n\u003cp\u003eGradient-weighted Class Activation Mapping (Grad-CAM) was applied to improve the interpretability of our 2.5D DL framework. This approach generates localization maps that emphasize image regions critical for the classification decision. By analyzing gradient flows in the final convolutional layer of the last residual block, we visualized activation patterns corresponding to the model\u0026rsquo;s predictions, thereby identifying salient regions that contributed most significantly to the classification outcome.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eCategorical and continuous variables were compared using the \u003cem\u003eχ\u0026sup2;\u003c/em\u003e test/Fisher's exact test and Mann-Whitney \u003cem\u003eU\u003c/em\u003e test/\u003cem\u003et\u003c/em\u003e-test, respectively. Model performance was evaluated by AUC, sensitivity, specificity, PPV, and NPV. A two-tailed \u003cem\u003eP\u003c/em\u003e-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 indicated significance. Analyses used R (v4.5.0) and Python 3.7.0 (with scikit-learn).\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003ePatient characteristics\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarizes the baseline characteristics of the 732 patients included in this retrospective study. The distribution of molecular subtypes and tumor grade showed no significant differences across the training, internal validation, and external validation sets (all \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026gt;\u0026thinsp;0.05). Luminal B was the predominant molecular subtype, and invasive ductal carcinoma was the most common histology. Significant differences between ALN-positive and ALN-negative patients were observed in MRI-reported ALN status and tumor size across all datasets (all \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), with positive MRI findings and larger tumor sizes being associated with ALN positivity. No significant differences were found in patient age or histology. While BI-RADS categories showed no significant differences in the training and internal validation sets, a significant disparity (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) was observed in the external validation set, with a higher prevalence of BI-RADS category 5 in the ALN-positive group.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePatient characteristics of three datasets\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"12\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c12\" colnum=\"12\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eCharacteristic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eTraining set (N\u0026thinsp;=\u0026thinsp;411)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c8\" namest=\"c6\"\u003e \u003cp\u003eInternal validation set (N\u0026thinsp;=\u0026thinsp;176)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c12\" namest=\"c10\"\u003e \u003cp\u003eExternal validation set (N\u0026thinsp;=\u0026thinsp;145)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eALN-negative\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;222)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eALN-positive\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;189)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eALN-negative\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;90)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eALN-positive\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;86)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c10\"\u003e \u003cp\u003eALN-negative\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;62)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c11\"\u003e \u003cp\u003eALN-positive\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;83)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c12\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMolecular_subtype\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.076\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.140\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e \u003cp\u003e0.389\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLuminal A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28 (13)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e15 (8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e15 (17)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e10 (12)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e13 (21)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e9 (11)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLuminal B\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e134 (60)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e137 (72)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e55 (61)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e6 (77)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e33 (53)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e51 (61)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHER 2+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e31 (14)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e20 (11)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e10 (11)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4 (5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e8 (13)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e13 (16)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTriple negative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e29 (13)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e17 (9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e10 (11)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e6 (7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e8 (13)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e10 (12)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGrade\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.173\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.274\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e \u003cp\u003e0.406\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e21 (9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e9 (5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e3 (5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e1 (1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e163 (73)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e149 (79)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e38 (61)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e56 (67)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e38 (17)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e31 (16)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e21 (34)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e26 (31)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMRI_report\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e193 (87)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e89 (47)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e82 (91)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e33 (38)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e61 (98)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e36 (43)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e29 (13)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e100 (53)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e8 (9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e53 (62)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e1 (2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e47 (84)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSize (mm)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026le;20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e93 (42)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e44 (23)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e44 (49)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e15 (17)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e39 (63)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e13 (16)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026gt;20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e129 (58)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e145 (77)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e46 (51)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e71 (83)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e23 (37)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e70 (84)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHisto_results\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.890\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.683\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e \u003cp\u003e0.693\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIDC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e215 (97)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e182 (96)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e85 (94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e84 (98)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e58 (94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e79 (95)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eILC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6 (3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6 (3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3 (3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2 (2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e3 (5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e4 (5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOthers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1 (0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1 (1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2 (2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e1 (2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, (years) Median (Q1, Q3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e50 (43, 57)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e50 (44, 54)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.258\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e51 (43.5, 56)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e51 (45, 56)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.878\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e53\u003c/p\u003e \u003cp\u003e(44, 56)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e50\u003c/p\u003e \u003cp\u003e(44, 56)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e \u003cp\u003e0.263\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBI_RADS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.211\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e104 (47)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e76 (40)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e50 (56)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e24 (28)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e42 (68)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e27 (33)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e118 (53)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e113 (60)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e40 (44)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e62 (72)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e20 (32)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e \u003cp\u003e56 (67)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"12\"\u003eNote: ALN, axillary lymph node; IDC, invasive ductal carcinoma; ILC, invasive lobular carcinoma; BI_RADS, Breast Imaging Reporting and Data System.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eFeature extraction and selection\u003c/h2\u003e \u003cp\u003eA total of 1817 radiomics features were extracted, along with 100 MIL_histogram, 100 MIL_TF-IDF, and 2048 2D DL features. Following feature selection with LASSO regression, 9 radiomics, 20 MIL_histogram, 24 MIL_TF-IDF, and 23 2D DL features were retained. The relative weights of the selected features are illustrated in \u003cb\u003eSupplementary Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/b\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eAnalysis of model performance\u003c/h2\u003e \u003cp\u003eThe MLP classifier was selected for subsequent analyses as it achieved the highest AUC on the internal validation set among all single-modality models across four feature types (\u003cb\u003eSupplementary Figure S2\u003c/b\u003e).\u003c/p\u003e \u003cp\u003eDiagnostic performance of all models is summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. In the training set, the stacking model achieved the highest AUC (0.962; 95% CI: 0.946\u0026ndash;0.977), with no significant difference compared to the MIL_histogram model (\u003cem\u003eP\u0026thinsp;=\u003c/em\u003e\u0026thinsp;0.103) but outperforming the MIL_TF‑IDF model (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.017). The MIL_histogram (AUC\u0026thinsp;=\u0026thinsp;0.955) and MIL_TF‑IDF (AUC\u0026thinsp;=\u0026thinsp;0.949) models performed similarly (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.137), both showing high specificity and PPV. In contrast, the Radiomics (AUC\u0026thinsp;=\u0026thinsp;0.778) and 2D DL (AUC\u0026thinsp;=\u0026thinsp;0.870) models were significantly inferior to the other three models (all \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePredictive Model Performance Metrics in the Datasets of This Study\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUC (95%CI)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePPV\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNPV\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTraining set\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMIL_histogram\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.955 (0.938\u0026ndash;0.972)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.831\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.941\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.924\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.867\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMIL_TF-IDF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.949 (0.931\u0026ndash;0.968)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.815\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.946\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.928\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.857\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRadiomics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.778 (0.735\u0026ndash;0.822)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.868\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.559\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.626\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.832\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2D DL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.870 (0.837\u0026ndash;0.903)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.841\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.730\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.726\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.844\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStacking\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.962 (0.946\u0026ndash;0.977)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.884\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.905\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.888\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.901\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eInternal validation set\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMIL_histogram\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.853 (0.798\u0026ndash;0.909)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.686\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.944\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.922\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.759\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMIL_TF-IDF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.849 (0.792\u0026ndash;0.905)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.721\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.900\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.873\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.771\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRadiomics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.737 (0.663\u0026ndash;0.811)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.651\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.733\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.700\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.687\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2D DL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.755 (0.683\u0026ndash;0.827)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.791\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.656\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.687\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.766\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStacking\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.885 (0.837\u0026ndash;0.933)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.628\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.738\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eExternal validation set\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMIL_histogram\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.855 (0.794\u0026ndash;0.915)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.747\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.887\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.899\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.724\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMIL_TF-IDF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.871 (0.812\u0026ndash;0.931)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.843\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.823\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.864\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.797\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRadiomics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.772 (0.695\u0026ndash;0.849)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.615\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.887\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.879\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.632\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2D DL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.778 (0.700-0.856)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.868\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.629\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.758\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.780\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStacking\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.890 (0.840\u0026ndash;0.939)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.639\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.968\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.964\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.667\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003eNote: AUC, area under the receiver operating characteristic curve; CI, confidence interval; PPV, positive predictive value; NPV, negative predictive value; MIL, multi-instance learning; TF-IDF, Term Frequency Inverse Document Frequency; 2D DL, two-dimensional deep learning.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn the internal validation set, the stacking model again showed the best performance (AUC\u0026thinsp;=\u0026thinsp;0.885; 95% CI: 0.837\u0026ndash;0.933), with perfect specificity and PPV (1.000), and significantly outperformed both MIL_histogram (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.013) and MIL_TF-IDF (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.009). The two MIL models remained comparable (P\u0026thinsp;=\u0026thinsp;0.694), while Radiomics (AUC\u0026thinsp;=\u0026thinsp;0.737) and 2D DL (AUC\u0026thinsp;=\u0026thinsp;0.755) were significantly worse than the top models (all \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/p\u003e \u003cp\u003eIn the external validation set, the stacking model achieved the highest AUC (0.890; 95% CI: 0.840\u0026ndash;0.939), along with high specificity (0.968) and PPV (0.964). It performed significantly better than the MIL_histogram model (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.014) but not the MIL_TF-IDF model (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.241). Radiomics and 2D DL models again demonstrated lower performance.\u003c/p\u003e \u003cp\u003eCalibration curves indicated that the stacking model was well-calibrated (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), and decision curve analysis revealed a superior net clinical benefit for this model (\u003cb\u003eSupplementary Figure S3\u003c/b\u003e). Representative clinical cases are shown in \u003cb\u003eSupplementary Figures S4-S5\u003c/b\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eSubgroup-based model performance assessment\u003c/h2\u003e \u003cp\u003eSubgroup analyses were conducted to assess the robustness of the predictive models based on factors previously associated with ALN metastasis, including patient age (\u0026le;\u0026thinsp;50 vs. \u0026gt;50 years), tumor size (\u0026le;\u0026thinsp;20 vs. \u0026gt;20 mm), histological grade (\u0026le;\u0026thinsp;2 vs. \u0026gt;2), and BI-RADS category (4 vs. 5) [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. The results, detailed in \u003cb\u003eSupplementary Figures S6-S9\u003c/b\u003e, revealed that most factors (age, tumor size, and BI-RADS category) did not significantly affect the performance of any model. Notably, histological grade had a significant impact on the performance of the 2D DL model. In contrast, the radiomics, MIL_histogram, MIL_TF-IDF, and stacking models all maintained consistent and robust performance across all subgroups.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis multi-center retrospective study developed and validated models based on 2.5D DL-MIL, 2D DL, and radiomics for the preoperative prediction of ALN metastasis in BC patients using the second phase of DCE-MRI. A stacking ensemble model was then constructed using the optimal single-modality models. The results demonstrated that the stacking model achieved the highest diagnostic performance, with AUCs of 0.962 (95% CI: 0.946\u0026ndash;0.977), 0.885 (95% CI: 0.837\u0026ndash;0.933), and 0.890 (95% CI: 0.840\u0026ndash;0.939) in the training, internal validation, and external validation sets, respectively. It also exhibited robust generalizability, satisfactory calibration, and a substantial net clinical benefit. Subgroup analyses confirmed the model's stable performance across patient subgroups stratified by age, tumor size, nuclear grade, and BI-RADS category.\u003c/p\u003e \u003cp\u003ePreoperative prediction of occult axillary lymph node (ALN) metastasis in breast cancer remains a significant clinical challenge. The Memorial Sloan Kettering Cancer Center (MSKCC) nomogram, the most widely used predictive tool, integrates clinicopathological parameters such as tumor size, lymphovascular invasion, and sentinel lymph node status to estimate non-sentinel lymph node metastasis risk [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. However, it lacks integration of quantitative imaging predictors. Although MRI features of primary breast cancer have demonstrated association with ALN metastasis and offer potential for improved preoperative prediction [\u003cspan additionalcitationids=\"CR24\" citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], their qualitative assessment is operator-dependent and fails to comprehensively capture imaging information. Recent radiomics approaches have attempted to address this limitation [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], but feature inconsistency across models hinders reproducibility. This generalizability concern is evidenced in our external validation, where the radiomics model achieved an AUC of only 0.772 (95% CI: 0.695\u0026ndash;0.849).\u003c/p\u003e \u003cp\u003eDeep learning (DL) approaches for lymph node assessment have been applied in several solid tumors. Multiple studies have investigated DL models for predicting ALN metastasis in BC patients [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. For example, Chen et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] developed a DCE-MRI-based CNN model to predict SLN and non-SLN metastasis in breast cancer, reporting AUCs of 0.899 (internal validation), 0.885 (external test 1), and 0.768 (external test 2) for SLN prediction, and 0.800, 0.763, and 0.728 for non-SLN prediction in the same cohorts. Guo et al. [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] developed multimodal models combining radiomics and DL features from mammography and MRI, with their combined MLP model achieving the highest test AUC of 0.846. However, most existing studies rely on 2D or 3D DL architectures. In our study, we proposed two 2.5D DL-MIL models based on DCE-MRI images of the primary tumor. Both models achieved promising and consistent performance, with AUCs ranging from 0.849 to 0.871 across the validation sets.\u003c/p\u003e \u003cp\u003eOur findings confirm the superiority of the 2.5D DL-MIL approach over the 2D DL model, aligning with previous studies [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. This improvement stems from several factors. Unlike 2D models relying on a single cross-sectional slice, the 2.5D DL-MIL model incorporates partial 3D context by aggregating multi-slice features across three anatomical planes using MIL. This multi-planar sampling preserves key spatial relationships between the tumor and its microenvironment\u0026mdash;features associated with ALN metastasis, such as peritumoral angiogenesis. Moreover, MIL reduces inter-slice variability and selection bias by treating each patient as a \u0026ldquo;bag\u0026rdquo; of instances, yielding a robust patient-level representation less affected by slice-specific noise.\u003c/p\u003e \u003cp\u003eIn contrast, the radiomics model showed the lowest performance (external validation AUC\u0026thinsp;=\u0026thinsp;0.772). Handcrafted radiomics features, based on predefined mathematical descriptors, often fail to capture complex, nonlinear tumor heterogeneity patterns linked to ALN metastasis. Deep learning models, however, automatically learn discriminative high-level features from imaging data, capturing subtle patterns beyond conventional radiomics and human perception.\u003c/p\u003e \u003cp\u003eThe high specificity of the stacking model (1.000 internally, 0.968 externally) holds significant clinical promise. Given that SLNB is associated with complications like arm numbness and lymphedema [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], a model that minimizes false positives is crucial. In our external validation, 96.8% of patients predicted as ALN-negative were pathologically negative, suggesting that only 3.2% might undergo unnecessary invasive staging. This aligns with the de-escalation goals of trials like ACOSOG Z0011 [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Additionally, the model's consistent performance across subgroups, including challenging cases like small tumors (\u0026le;\u0026thinsp;20 mm) where it maintained an AUC\u0026thinsp;\u0026gt;\u0026thinsp;0.870, addresses a key limitation of many existing prediction tools.\u003c/p\u003e \u003cp\u003eSeveral limitations of this study warrant consideration. First, its retrospective nature may introduce selection bias; prospective validation in a consecutive cohort is needed to confirm real-world clinical utility. Second, the model was trained exclusively on the second post-contrast DCE-MRI phase; integrating additional sequences (e.g., DWI) might improve performance. Third, the external validation was performed with data from a single additional center; future multi-center studies with larger, more diverse cohorts (e.g., including 1.5-T MRI data) are necessary to enhance generalizability. Finally, although ROIs were delineated by experienced radiologists and reviewed to minimize variability, inter-observer differences in segmentation could influence feature extraction. Future work could explore automated segmentation to standardize this process.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThe 2.5D DL-MIL-based stacking model provides a non-invasive, accurate, and robust tool for the preoperative prediction of ALN metastasis in BC. By integrating complementary features from multiple approaches, it outperforms single-modality models and demonstrates consistent efficacy across diverse patient subgroups. Its high specificity has the potential to reduce unnecessary invasive procedures, supporting its broader clinical adoption. Future prospective, multi-center studies integrating multi-parametric MRI data are warranted to further validate and refine this tool for personalized treatment planning.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eALN\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eAxillary lymph node\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eALND\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eAxillary lymph node dissection\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSLNB\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSentinel lymph node biopsy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDCE-MRI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDynamic contrast-enhanced magnetic resonance imaging\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eROI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eRegion of interest\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e2D/3D\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eTwo-dimensional/three-dimensional\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDL\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDeep learning\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMIL\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMulti-instance learning\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eBI-RADS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eBreast Imaging Reporting and Data System\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAUC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eArea under the curve\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eICC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eIntraclass correlation coefficient\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLASSO\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eLeast absolute shrinkage and selection operator\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eROC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eReceiver operating characteristic\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDCA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDecision curve analysis.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":" \u003cp\u003e \u003cb\u003eEthics declaration\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThis retrospective study was conducted in accordance with the Declaration of Helsinki. The study was approved by the Ethics Committee of Shangqiu Medical College (approval number: SYLL2025-008). Given the use of anonymized data, informed consent was waived for this study. Clinical trial number: Not applicable.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent for publication\u003c/strong\u003e \u003cp\u003eAs all patient data utilized in this retrospective study were fully anonymized and could not be traced back to individual participants, informed consent for publication was formally waived.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eCompeting interests\u003c/strong\u003e \u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThis work was supported by the Science and Technology Research Project of Henan Province (252102310061), Key Scientific Research Projects of Higher Education Institutions of Henan Province (25B320031, 26B320017), Science and Technology Program of Shangqiu City (2024075, 2025068), and Doctoral Scientific Research Initiation Project of Shangqiu Medical College (BSJH004). Additional support was provided by the Henan Provincial Health Commission Key Laboratory of Imaging Precision Diagnosis.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eLSM: Data curation, Writing \u0026ndash; Original draft preparation. XZ: Conceptualization, Supervision, Visualization. YXZ: Conceptualization, Methodology, Supervision. LL: Conceptualization, Supervision, Visualization. XM: Conceptualization, Writing \u0026ndash; Review \u0026amp; Editing. SYL: Resources, Formal analysis. FMS: Visualization, Formal analysis. XAZ: Conceptualization, Visualization, Writing \u0026ndash; Review \u0026amp; Editing. All authors have read and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets employed and/or examined in this study can be obtained from the corresponding author upon reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSiegel RL, Miller KD, Wagle NS, Jemal A. Cancer statistics, 2023. CA: A Cancer. J Clin. 2023;73:17\u0026ndash;48.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeSantis CE, Ma J, Gaudet MM, Newman LA, Miller KD, Goding Sauer A, et al. Breast cancer statistics, 2019. CA: A Cancer. J Clin. 2019;69:438\u0026ndash;51.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSiegel RL, Miller KD, Jemal A, Cancer statistics. 2017. CA: A Cancer Journal for Clinicians. 2017;67:7\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYu Y, Tan Y, Xie C, Hu Q, Ouyang J, Chen Y, et al. Development and Validation of a Preoperative Magnetic Resonance Imaging Radiomics-Based Signature to Predict Axillary Lymph Node Metastasis and Disease-Free Survival in Patients With Early-Stage Breast Cancer. JAMA Netw open. 2020;3:e2028086.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGentilini O, Veronesi U. Abandoning sentinel lymph node biopsy in early breast cancer? A new trial in progress at the European Institute of Oncology of Milan (SOUND: Sentinel node vs Observation after axillary UltraSouND). Breast (Edinburgh, Scotland). 2012;21:678\u0026ndash;681.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiang Y, Chen X, Tong Y, Zhan W, Zhu Y, Wu J, et al. Higher axillary lymph node metastasis burden in breast cancer patients with positive preoperative node biopsy: may not be appropriate to receive sentinel lymph node biopsy in the post-ACOSOG Z0011 trial era. World J Surg Oncol. 2019;17:37.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoughey JC, Moriarty JP, Degnim AC, Gregg MS, Egginton JS, Long KH. Cost modeling of preoperative axillary ultrasound and fine-needle aspiration to guide surgery for invasive breast cancer. Ann Surg Oncol. 2010;17:953\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLanger I, Guller U, Berclaz G, Koechli OR, Schaer G, Fehr MK, 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\u0026ndash;61.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLe Boulc'h M, Gilhodes J, Steinmeyer Z, Moli\u0026egrave;re S, Mathelin C. Pretherapeutic Imaging for Axillary Staging in Breast Cancer: A Systematic Review and Meta-Analysis of Ultrasound, MRI and FDG PET. J Clin Med. 2021;10(7):1543.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYu Y, He Z, Ouyang J, Tan Y, Chen Y, Gu Y, 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.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi X, Yang L, Jiao X. Comparison of Traditional Radiomics, Deep Learning Radiomics and Fusion Methods for Axillary Lymph Node Metastasis Prediction in Breast Cancer. Acad Radiol. 2023;30:1281\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuo YJ, Yin R, Zhang Q, Han JQ, Dou ZX, Wang PB, et al. MRI-Based Kinetic Heterogeneity Evaluation in the Accurate Access of Axillary Lymph Node Status in Breast Cancer Using a Hybrid CNN‐RNN Model. J Magn Reson Imaging. 2024;60:1352\u0026ndash;64.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y, Liao Q, Ding L, Zhang J. Bridging 2D and 3D segmentation networks for computation-efficient volumetric medical image segmentation: An empirical study of 2.5D solutions. Comput Med Imaging Graph. 2022;99:102088.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZeng Y, Zhang X, Kawasumi Y, Usui A, Ichiji K, Funayama M, et al. A 2.5D Deep Learning-Based Method for Drowning Diagnosis Using Post-Mortem Computed Tomography. IEEE J biomedical health Inf. 2023;27:1026\u0026ndash;35.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y-B, Chen Z-Q, Bu Y, Lei P, Yang W, Zhang W. Construction of a 2.5D Deep Learning Model for Predicting Early Postoperative Recurrence of Hepatocellular Carcinoma Using Multi-View and Multi-Phase CT Images. J Hepatocellular Carcinoma. 2024;11:2223\u0026ndash;39.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu J, Zou L, Xie X, Xu R, Tian Y, Zhang B. 2.5D deep learning based on multi-parameter MRI to differentiate primary lung cancer pathological subtypes in patients with brain metastases. Eur J Radiol. 2024;180:111712.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang Z, Jiang H, Shan S, Wang X, Kou Q, Wang C, et al. 2.5D Deep Learning-Based Prediction of Pathological Grading of Clear Cell Renal Cell Carcinoma Using Contrast-Enhanced CT: A Multicenter Study. Acad Radiol. 2025;32:5907\u0026ndash;16.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe J, Xu J, Chen W, Cao M, Zhang J, Yang Q, et al. Development of a deep learning model for T1N0 gastric cancer diagnosis using 2.5D radiomic data in preoperative CT images. npj Precision Oncol. 2025;9(1):249.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Z, Hong M, Li X, Lin L, Tan X, Liu Y. Predicting axillary lymph node metastasis in breast cancer patients: A radiomics-based multicenter approach with interpretability analysis. Eur J Radiol. 2024;176:111522.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu H, Zou L, Xu N, Shen H, Zhang Y, Wan P, et al. Deep learning radiomics based prediction of axillary lymph node metastasis in breast cancer. npj Breast Cancer. 2024;10(1):22.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen M, Kong C, Lin G, Chen W, Guo X, Chen Y, et al. Development and validation of convolutional neural network-based model to predict the risk of sentinel or non-sentinel lymph node metastasis in patients with breast cancer: a machine learning study. eClinicalMedicine. 2023;63:102176.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMemorial Sloan Kettering Cancer Center breast cancer nomogram. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://nomo grams.mskcc.org/breast/index.aspx\u003c/span\u003e\u003cspan address=\"http://nomo grams.mskcc.org/breast/index.aspx\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Accessed June 18, 2019.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLuo H, Chen Z, Xu H, Ren J, Zhou P. Peritumoral edema enhances MRI-based deep learning radiomic model for axillary lymph node metastasis burden prediction in breast cancer. Sci Rep. 2024;14:18900.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaltzer PA, Dietzel M, Burmeister HP, Zoubi R, Gajda M, Camara O, et al. Application of MR mammography beyond local staging: is there a potential to accurately assess axillary lymph nodes? evaluation of an extended protocol in an initial prospective study. AJR Am J Roentgenol. 2011;196:W641\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIima M, Honda M, Sigmund EE, Ohno Kishimoto A, Kataoka M, Togashi K. Diffusion MRI of the breast: Current status and future directions. J Magn Reson Imaging. 2020;52:70\u0026ndash;90.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHan L, Zhu Y, Liu Z, Yu T, He C, Jiang W, et al. Radiomic nomogram for prediction of axillary lymph node metastasis in breast cancer. Eur Radiol. 2019;29:3820\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuo F, Sun S, Deng X, Wang Y, Yao W, Yue P, et al. Predicting axillary lymph node metastasis in breast cancer using a multimodal radiomics and deep learning model. Front Immunol. 2024;15:1482020.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMai W, Fan X, Zhang L, Li J, Chen L, Hua X, et al. Deep learning for differential diagnosis of parotid tumors based on 2.5D magnetic resonance imaging. Ann Med. 2025;57(1):2520401.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGiuliano AE, Ballman KV, McCall L, Beitsch PD, Brennan MB, Kelemen PR, 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\u0026ndash;26.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDescription. of each illustration.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Breast cancer, Axillary lymph node metastasis, Deep learning, Magnetic resonance imaging, Multi-instance learning","lastPublishedDoi":"10.21203/rs.3.rs-8372278/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8372278/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eObjective\u003c/h2\u003e \u003cp\u003eTo compare the performance of 2.5D deep learning (DL) with multi-instance learning (MIL), 2D DL, and radiomics models in predicting axillary lymph node (ALN) metastasis in breast cancer (BC) patients using dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI).\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eIn this study, 732 patients from two independent institutions who underwent preoperative DCE-MRI were included. Based on the primary tumor region, we developed and compared four single-modality prediction models: a radiomics model, a 2D DL model, and two 2.5D DL-MIL models using different feature aggregation strategies. A stacking model was subsequently constructed by integrating the optimal single-modality models. The models\u0026rsquo; performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and Decision Curve Analysis.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe stacking model achieved the highest predictive performance, with AUCs of 0.962 (95% CI: 0.946\u0026ndash;0.977) in the training set, 0.885 (95% CI: 0.837\u0026ndash;0.933) in the internal validation set, and 0.890 (95% CI: 0.840\u0026ndash;0.939) in the external validation set. It demonstrated high specificity (1.000 and 0.968 in the internal and external validation sets, respectively) and provided a significant net clinical benefit. The 2.5D DL-MIL models significantly outperformed the 2D DL and radiomics models. Furthermore, the stacking model maintained robust performance across key clinical subgroups defined by age, tumor size, and BI-RADS category.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThe 2.5D DL-MIL-based stacking model provides an accurate, non-invasive tool for preoperative prediction of ALN metastasis in BC. Its high specificity holds promise for reducing unnecessary sentinel lymph node biopsies, thereby aiding in personalized surgical planning.\u003c/p\u003e","manuscriptTitle":"Comparative Analysis of 2.5D Deep Learning, 2D Deep Learning, and Radiomics Models for Predicting Axillary Lymph Node Metastasis in Breast Cancer: A Multi-Center Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-19 13:41:24","doi":"10.21203/rs.3.rs-8372278/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-01-21T14:28:17+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"284742760408122426001484807405667659923","date":"2026-01-14T16:37:12+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-01-14T15:18:44+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-13T04:06:02+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-12-19T04:40:03+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-12-18T16:29:06+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2025-12-18T15:06:35+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"c069c9fe-3c61-4638-8cf1-393a00d0dc34","owner":[],"postedDate":"January 19th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-01-19T13:41:24+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-19 13:41:24","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8372278","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8372278","identity":"rs-8372278","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}