Multiparametric MRI-based radiomics combined with machine learning for preoperative differentiation of luminal A and luminal B breast cancer: a multicenter study

Multiparametric MRI-based radiomics combined with machine learning for preoperative differentiation of luminal A and luminal B breast cancer: a multicenter 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 Multiparametric MRI-based radiomics combined with machine learning for preoperative differentiation of luminal A and luminal B breast cancer: a multicenter study Hui zhou, Daoyu Yang, Shiguang Li, Xudong Liu, Ling Wei, Jiarui Wang, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8793807/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 12 You are reading this latest preprint version Abstract Background Accurate preoperative differentiation of luminal A from luminal B breast cancer is critical for individualized treatment planning, yet current assessment relies on invasive biopsy with inherent sampling limitations. This study aimed to develop and validate a multiparametric MRI-based radiomics model for non-invasive luminal subtype classification across independent institutions. Methods This retrospective multicenter study included 130 patients with pathologically confirmed luminal A (n=44) or luminal B (n=86) invasive ductal carcinoma from three centers. Radiomics features were extracted from T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and dynamic contrast-enhanced MRI (DCE-MRI) using PyRadiomics. A four-stage feature selection pipeline (variance thresholding, Pearson correlation, ANOVA F-test, LASSO regression) was applied for dimensionality reduction. Synthetic Minority Over-sampling Technique (SMOTE) was used within training folds to address class imbalance. Nine machine learning classifiers were systematically compared. The model was developed using Center 1 data (n=64) with five-fold cross-validation and externally validated in two independent cohorts [Center 2 (n=36) and Center 3 (n=30)]. Performance was evaluated using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and decision curve analysis (DCA). Bootstrap resampling (1,000 iterations) assessed metric stability. Results Among nine classifiers, XGBoost demonstrated the best performance. The clinical-radiomics fusion model achieved AUCs of 0.939 (95% CI: 0.868–0.986), 0.897 (95% CI: 0.706–1.000), and 0.870 (95% CI: 0.708–0.986) in Centers 1, 2, and 3, respectively. This model significantly outperformed single-sequence approaches in pairwise DeLong tests (P<0.05) and demonstrated superior net benefit in DCA. SHAP analysis revealed that Ki-67 index, progesterone receptor status, patient age, and radiomics features from DWI and DCE-MRI were the most influential predictors, with consistent importance rankings across all three centers. Conclusions The multiparametric MRI radiomics model integrating T2WI, DWI, and DCE-MRI demonstrates robust and generalizable performance for differentiating luminal A from luminal B breast cancer across multiple independent centers. SHAP-based interpretability analysis enhances clinical transparency by identifying consistent predictive features. The model offers potential clinical utility as a non-invasive tool when biopsy is contraindicated and as a supplement to immunohistochemistry by providing whole-tumor heterogeneity assessment. Breast cancer Magnetic resonance imaging Radiomics Machine learning Luminal subtype XGBoost SHAP Multicenter validation Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction According to 2022 global cancer statistics, ~ 2.3 million women worldwide were diagnosed with breast cancer, underscoring its status as a major public health concern( 1 ). The Prediction Analysis of Microarray 50 (PAM50) gene expression assay categorizes breast cancer tumors into five intrinsic subtypes-luminal A, luminal B, human epidermal growth factor receptor 2 (HER2)-enriched, basal-like and normal-like-based on the messenger ribonucleic acid (mRNA) expression of 50 genes ( 2 – 4 ). Among these, luminal A and B are the most common subtypes ( 5 ), but they differ markedly in treatment approaches and prognosis. Despite sharing hormone receptor positivity, luminal B tumors exhibit higher proliferative activity and lower hormone receptor expression, resulting in worse prognosis and higher recurrence rates compared with luminal A ( 6 – 8 ). Accurate preoperative differentiation between these two subtypes is crucial for tailoring treatment and improving patient outcomes. However, current classification methods, although effective, often rely on invasive, costly, and time-consuming procedures, such as histopathological biopsy and molecular profiling ( 9 , 10 ). Magnetic resonance imaging (MRI) techniques such as T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI) and dynamic contrast-enhanced MRI (DCE-MRI) have emerged as promising non-invasive tools for breast tumor assessment. These modalities offer complementary information, with T2WI reflecting anatomical structures and tissue contrast, DWI providing insights into cellular density and microstructural integrity and DCE-MRI capturing perfusion characteristics and vascular dynamics ( 11 , 12 ). Nevertheless, their standalone diagnostic performance remains limited, particularly in capturing tumor heterogeneity and reliably distinguishing between luminal A and luminal B subtypes ( 13 ). This limitation underscores the need for advanced analytical approaches to integrate multiparametric MRI data to enhance diagnostic precision and facilitate personalized treatment decision-making. Radiomics, introduced by Lambin( 14 ), has emerged as a valuable tool for extracting high-throughput quantitative features from medical images, converting them into data that can be analyzed for tumor heterogeneity ( 15 – 17 ). By converting medical images into structured data, radiomics facilitates objective tumor characterization beyond visual interpretation. When combined with machine learning algorithms such as support vector machines and random forests, radiomics has shown potential in tumor classification and outcome prediction ( 18 – 21 ). Despite its success in various oncologic contexts, studies specifically addressing the classification of luminal A and B breast cancer subtypes remain limited. These two subtypes exhibit overlapping imaging phenotypes but differ notably in molecular expression profiles, treatment response and prognostic outcomes, posing a considerable diagnostic challenge. Accurate non-invasive differentiation between these subtypes is crucial for guiding personalized treatment decisions, particularly when biopsy is contraindicated or with insufficient tissue sampling. Although radiomics-based approaches have been explored for breast cancer molecular subtyping, several gaps remain. First, most studies have focused on distinguishing triple-negative or HER2-enriched subtypes from others, few studies have specifically addressed the clinically challenging task of differentiating luminal A from luminal B subtypes ( 21 , 22 ). Second, the majority of existing studies were conducted at single centers using single MRI sequences, limiting the generalizability of the findings( 23 ). Third, systematic comparisons of multiple machine learning classifiers using identical feature sets are rarely reported, making it difficult to identify the optimal modeling strategy( 24 ). Fourth, although explainability techniques such as SHAP have been applied in breast cancer imaging, their integration with multiparametric MRI radiomics and multicenter validation for luminal subtype classification has not been explored( 25 ). Therefore, the present study aimed to develop and validate a radiomics-based multiparametric model that integrates MRI-derived imaging features, clinical characteristics, pathological markers [estrogen receptor (ER), progesterone receptor (PR), HER2, Ki-67] and lymph node status to distinguish luminal A from luminal B breast cancers. By leveraging multi-center datasets and adopting an integrated radio-genomic approach, the present study aimed to enhance the robustness, accuracy and generalizability of preoperative subtype classification to facilitate personalized treatment planning. Materials and methods Data acquisition . This retrospective multicenter study was conducted at three institutions (Center 1, Center 2, and Center 3). A total of 136 anonymized breast cancer cases diagnosed between October 2020 and June 2024 were included. All patients underwent preoperative MRI, including T2WI, DWI and DCE-MRI. The detailed scanning protocols are provided in Supplement Table S1. After excluding patients with incomplete imaging data, 130 patients with pathologically confirmed luminal A or luminal B invasive ductal carcinoma and complete multimodal MRI were included in the final analysis. The inclusion criteria were as follows: (i) Age 18-80 years; (ii) histopathologically confirmed luminal A or B invasive ductal carcinoma; (iii) complete preoperative breast MRI including T2WI, DWI and DCE-MRI; (iv) available clinicopathological data (ER, PR, HER2, Ki-67); and (v) treatment-naive status. The exclusion criteria were as follows: (i) Patients with incomplete imaging data across any of the three MRI modalities; (ii) poor image quality affecting feature extraction; (iii) previous breast cancer treatment; and (iv) missing critical pathological markers. Subsequently, data from Center 1, comprising 64 patients, were used for model development and internal evaluation using five-fold cross-validation. Data from the remaining two institutions, Center 2 (n=36) and Center 3 (n=30), served as independent external validation cohorts. Fig. 1 illustrates the flowchart detailing the inclusion and exclusion process for the study cohort. Image segmentation and radiomics feature extraction. In the present study, tumor segmentation was performed manually on a slice-by-slice basis (2D approach) across three imaging modalities (DCE-MRI, T2WI, DWI) using ITK-SNAP software (version 3.8.0; itksnap.org) (Fig. 2). For each case, the tumor was delineated on all axial slices where the lesion was visible, and the individual 2D contours were subsequently combined to generate a three-dimensional volume of interest (ROI) for radiomics feature extraction. All delineations were subsequently reviewed and verified by a senior radiologist with over 20 years of experience to ensure anatomical accuracy and consistency.Radiomics features were extracted from these ROIs using the PyRadiomics package (version 3.0; pyradiomics.readthedocs.io), covering a wide range of descriptors: First-order statistics reflecting intensity distribution; shape-based features characterizing tumor geometry; and texture features derived from gray-level co-occurrence matrix, gray-level run length matrix, gray-level size zone matrix (GLSZM), gray-level dependence matrix and neighboring gray-tone difference matrix. Prior to feature extraction, all MRI images were resampled to an isotropic voxel spacing of 1 × 1 × 1 mm³ using B-spline interpolation to minimize the influence of voxel size heterogeneity across centers. Voxel intensities were subsequently standardized using z-score normalization within the tumor volume (mean = 0, standard deviation = 1). Gray-level discretization was then performed using a fixed bin width of 25 to ensure consistent texture quantization across all cases. To improve tumor heterogeneity and spatial complexity sensitivity, image filters including Laplacian of Gaussian filters with s values of 1.0, 2.0 and 3.0 and multi-level wavelet decompositions were applied before feature extraction. All extracted radiomics features were further standardized using z-score scaling before model construction to improve feature comparability across centers. Features selection and model construction . A multi-step feature selection procedure was employed to optimize model input, with all steps performed strictly within the training folds of five-fold cross-validation. Initially, features exhibiting low variance were excluded using a variance threshold method (threshold, 0.01) to remove uninformative variables. Subsequently, highly correlated features were identified using pairwise Pearson correlation analysis. For each feature pair with a correlation coefficient greater than 0.9, the feature with the lower univariate analysis of variance (ANOVA) F-value was removed while the more informative feature was retained, thereby mitigating redundancy and multicollinearity. The remaining feature set was refined through univariate analysis employing the ANOVA F-test, selecting the top features (up to 100) based on their statistical significance. The least absolute shrinkage and selection operator (LASSO) with five-fold cross-validation was subsequently applied to select the optimal subset of features to enhance sparsity and interpretability. Furthermore, clinical features including demographic, pathological and imaging-derived variables were integrated with the final selected radiomics features to develop a comprehensive clinical-radiomics model, aiming to exploit the complementary strengths of both data types for enhanced breast cancer subtype classification. Machine learning model development. To address the class imbalance inherent in the dataset, the Synthetic Minority Over-sampling Technique from the “imbalanced-learn” library (version 0.12.4; https://imbalanced-learn.org/) was applied only within the training folds of each cross-validation split (26). Due to the substantial variability in model performance introduced by different machine learning algorithms, a comprehensive set of machine learning classifiers was subsequently evaluated on the combined clinical-radiomics features, including logistic regression (27), k-nearest neighbors (KNN) (28), naive Bayes (29), support vector machine (SVM) (30), decision tree (31), random forest (32), Adaptive Boosting (AdaBoost) (33), gradient boosting decision tree (GBDT) (34) and Extreme Gradient Boosting (XGBoost) (35). Model development was performed using five-fold cross-validation, with the best-performing fold selected for each model based on validation metrics. All models were trained using default hyperparameters (Supplement Table 2). The models chosen throughout this procedure were subsequently externally validated on independent cohorts from Centers 2 and 3 to assess their generalizability and predictive performance. Model evaluation and statistical analysis . The performance of the radiomics models was comprehensively evaluated using several metrics, including the area under the curve (AUC), sensitivity, specificity and accuracy. Receiver operating characteristic (ROC) analysis was performed to assess the discriminatory ability of each model. To facilitate model interpretability, SHapley Additive exPlanations (SHAP) values were calculated to quantify the contribution of each feature to model predictions. Comparative analysis of different classifiers was performed using DeLong’s test to determine the statistical significance of differences in AUCs. Clinical utility of each model was further assessed using decision curve analysis (DCA) by calculating the net benefit across a range of threshold probabilities. To ensure the reliability of the research findings, bootstrapping with 1,000 iterations was employed to estimate confidence intervals (CI) for performance metrics. For continuous variables, normality was assessed using the Shapiro–Wilk test implemented in Python (SciPy library, version 1.31.1) (Supplement Table S3). Group comparisons were performed using unpaired Student’s t-test when the data conformed to a normal distribution, and the non-parametric Mann-Whitney U test was applied when normality was not met. Categorical variables were analyzed using the Fisher’s exact test. P<0.05 was considered to indicate a statistically significant difference. Results Patient characteristics . The clinical and molecular characteristics of patients with luminal A and luminal B breast cancer across the three independent centers are summarized in Table 1. In the cohort from Center 1, patients with luminal A were significantly older compared with those with the luminal B subtype (mean age, 57.05 vs. 49.00 years; P=0.008). ER and PR levels were also significantly higher in the luminal A group (ER, 79.73 vs. 67.00%; P=0.025; PR, 66.05 vs. 27.07%; P<0.001), whereas Ki-67 expression was significantly lower (16.58 vs. 34.31%; P0.05). In Center 2, Ki-67 was the only variable showing a statistically significant difference between subtypes, being markedly lower in luminal A tumors compared with luminal B tumors (6.86 vs. 35.00%; P<0.001). Other clinical and pathological variables, including age, ER and PR expression, tumor margins, morphology, HER-2 status and lymph node status, did not differ significantly between luminal A and B subtypes (all P>0.05). Data from Center 3 also revealed significant differences in PR and ER levels (PR, 76.67 vs. 28.75%; P＜0.001; ER, 84.44 vs. 51.67%; P=0.018) in the luminal A group compared with the luminal B group. HER-2 status distribution also varied significantly between the subtypes (P=0.027). Other characteristics, including age, Ki-67, tumor margins, morphology and lymph node status, did not show significant differences (all P>0.05). Feature extraction and selection . In the present study, 1,130 radiomic features were initially extracted from each imaging modality. After dimensionality reduction and feature selection, 24 features from DCE-MRI, 14 from T2WI and 16 from DWI were retained for constructing the respective single-parameter models. For the multiparametric fusion models, 14 features were selected for combining DCE-MRI and T2WI, 11 for DCE-MRI and DWI, 13 for T2WI and DWI and 14 for the integrative model incorporating all three MRI sequences. In addition, a clinical-radiomics fusion model was developed by integrating 10 clinical variables including age, sex, tumor size, histological type, tumor margins, tumor shape, ER, PR, lymph node status and Ki-67 index. SHAP-based interpretability analysis of the clinical-radiomics fusion model revealed that clinical and imaging features contributed substantially to subtype classification (Fig. 3, Supplement Fig. 2). Among the top-ranking predictors, Ki-67 index, PR status and patient age demonstrated the most significant influence on the model’s output, underscoring the critical role of clinical biomarkers. In parallel, radiomic features derived from DWI (original_shape_MeshVolume) and DCE-MRI (wavelet-LHL_glszm_LargeAreaEmphasis) also ranked highly, indicating that imaging phenotypes provided complementary predictive information. To assess the stability of feature importance across centers, SHAP analysis was extended to both external validation cohorts (Centers 2 and 3). Across all three cohorts, clinical variables (including Ki-67, PR status and age), together with several radiomics features (e.g., DWI_original_shape_MeshVolume), consistently exhibited high importance, indicating that the model relied on stable and generalizable predictors rather than cohort-specific patterns. Comparative evaluation of machine learning classifiers . The comparative performance of the machine learning classifiers across the three centers is depicted in Fig. 4. Classifiers such as KNN, naive Bayes, SVM, and logistic regression exhibited consistently suboptimal performance, with AUC values <0.70 in all cohorts, indicating limited discriminative capability. Although the decision tree classifier yielded the highest AUC in Center 1 (0.9450), its performance declined substantially in the external validation cohorts from Center 2 and Center 3 (AUCs of 0.7167 and 0.7083, respectively), indicating limited generalizability. By contrast, the random forest classifier demonstrated consistently robust performance across all cohorts, achieving AUCs of 0.8942, 0.8695 and 0.8935 in Centers 1, 2 and 3, respectively. The GBDT classifier also showed favorable performance in Center 1 and Center 3, with AUCs of 0.9427 and 0.8750, respectively; however, its discriminative ability declined in Center 2 (AUC =0.8128). Among all evaluated classifiers, XGBoost achieved the most stable and high-level performance across the three centers, with AUCs of 0.9392, 0.8966 and 0.8704, respectively. Pairwise AUC comparisons using DeLong’s test demonstrated that XGBoost outperformed most of the other classifiers across the three cohorts (Supplement Fig.1), with the majority of comparisons reaching statistical significance (P < 0.05), reflecting superior generalizability and predictive robustness. Consequently, XGBoost was selected as the optimal model for subsequent analyses in the present study. Classification performance of the radiomics models. Radiomics models based on single imaging modalities showed moderate classification performance, while multimodal fusion models markedly improved discrimination between luminal A and luminal B subtypes (Fig. 5). The clinical-radiomics fusion model consistently achieved the highest AUCs of 0.9392, 0.8966 and 0.8704 in Centers 1, 2 and 3, respectively, followed by the fusion of all imaging features (T2WI, DCE-MRI, DWI), which yielded AUCs of 0.9088, 0.8621 and 0.8426. By contrast, the T2WI-based model demonstrated the worst performance across all centers, with AUCs of 0.8018, 0.7340 and 0.7083, respectively. In addition, the classification metrics across the three centers consistently supported the superior performance of the clinical-radiomics fusion model, as detailed in Tables 2-4 (Center 1-3). In Center 1, the fusion model achieved a sensitivity of 0.8000, specificity of 0.9474, positive predictive value (PPV) of 0.9730 and negative predictive value (NPV) of 0.6667, alongside an AUC of 0.9392. Similar results were observed in Centers 2 and 3, with accuracies of 0.9167 (PPV, 0.9643) and 0.7667 (PPV, 0.6667) and AUCs of 0.8966 and 0.8704, respectively. Notably, including clinical variables markedly enhanced sensitivity and NPV, indicating improved identification of luminal B cases. Among models without clinical data, the fusion of all imaging features yielded the best results across all cohorts (AUCs, 0.9088, 0.8621, 0.8426), substantially outperforming unimodal models. By contrast, single-modality models, particularly those based on T2WI, showed consistently lower sensitivity, specificity and AUC values, with reduced generalizability. These results demonstrate that integrating radiomics features from multiple MRI sequences with clinical variables yields a reliable and generalizable model for subtype classification, supporting its potential applicability in real-world multi-center settings. DCA was performed across Centers 1 to 3 to further assess the predictive models’ clinical utility, as shown in Fig. 6. In Center 1, the clinical-radiomics model demonstrated the most favorable net benefit when the threshold probability >0.3, followed by the fusion model incorporating all imaging modalities, and the model combining DCE-MRI and DWI features. Models based on single imaging modalities exhibited limited net benefit, especially at higher thresholds. A similar trend was observed in Center 2, where the clinical-radiomics model maintained the most notable net benefit, while multimodal fusion models remained superior to unimodal approaches. In Center 3, although the overall net benefits declined, the clinical-radiomics model continued to outperform others, particularly within the low-to-intermediate threshold range (0.2-0.5). Single-modality models exhibited limited clinical value across most threshold probabilities, highlighting the superior effectiveness of the clinical-radiomics fusion model in enhancing decision-making and providing more reliable clinical guidance. Differences in model performance . To further evaluate the differences in discriminatory performance among the eight predictive models (Model 1, DCE-based radiomics, Model 2: T2WI-based radiomics, Model 3: DWI-based radiomics, Model 4: T2WI + DCE fusion, Model 5: T2WI + DWI fusion, Model 6: DCE + DWI fusion, Model 7: T2WI + DCE + DWI fusion, and Model 8: T2WI + DCE + DWI + Clinical fusion), pairwise DeLong tests were performed to compare AUCs across the three centers, as illustrated in Fig. 7. In Center 1, most comparisons (Model 1 vs. Models 4, 7, 8; Model 2 vs. Models 4, 6, 7, 8; Model 3 vs. Models 4, 8; and Model 4 vs. Models 5, 6, 7, 8) showed statistically significant differences (P<0.05), especially between single-modality and multi-modality fusion models. The clinical-radiomics fusion model demonstrated significantly higher AUCs compared with most other models. By contrast, differences between it and other multimodal fusion models were not statistically significant (P>0.05), indicating comparable performance among complex models. In Center 2, the clinical-radiomics fusion model consistently outperformed several single-modality and specific dual-modality models, demonstrating its enhanced discriminatory capability. However, the fusion model integrating all imaging modalities without clinical data did not show statistically significant improvements over single-modality models, indicating a diminished incremental benefit. Moreover, no significant differences were observed among the single-modality models, suggesting comparable diagnostic performance within this subgroup. In Center 3, most pairwise comparisons, particularly among fusion-based models, did not demonstrate statistical significance. The lack of discernible differences between the clinical-radiomics fusion model and other fusion approaches suggests a convergence in predictive performance within this cohort, thereby highlighting the inherent challenges in attaining consistent generalizability across heterogeneous patient populations. Discussion Radiomics quantitatively extracts high-dimensional features (texture, shape, gray-level statistics) from medical images, reflecting tumor microheterogeneity and molecular variations. In the present study, a multiparametric radiomics model was developed by integrating T2WI, DWI and DCE-MRI to differentiate luminal A from luminal B breast cancer subtypes. The fusion model achieved an AUC of 0.9392 in the training cohort and 0.8966 and 0.8704 in two independent test cohorts, significantly outperforming single-modality approaches. This performance was consistently maintained across three institutions despite variations in acquisition protocols, indicating good generalizability and clinical applicability. The diagnostic advantage of the fusion model can be attributed to the complementary roles of each MRI sequence. DCE-MRI quantifies tumor vascularity and perfusion, both closely associated with angiogenesis. Previous studies have demonstrated that luminal B tumors exhibit higher Ktrans (volume transfer constant) and kep (rate constant) values compared with luminal A tumors, reflecting more active angiogenesis (36, 37). DWI measures water diffusion restriction, with lower apparent diffusion coefficient values in luminal B tumors reflecting higher cellularity (38). T2WI depicts internal architecture and stromal composition, with luminal A lesions more often showing homogeneous signal intensity and abundant fibroglandular stroma (13). By integrating these modalities, the multiparametric clinical-radiomics model captures vascular, microstructural and compositional heterogeneity in a unified framework, likely contributing to its superior performance in Centers 1 and 2 and a robust, albeit slightly lower, performance in Center 3 (39). Recognizing that model generalizability depends not only on feature diversity but also on classifier choice, machine learning algorithms were systematically compared. XGBoost achieved optimal performance across the three centers (AUCs, 0.9392, 0.8966, 0.8704), outperforming traditional classifiers. Pairwise DeLong tests confirmed that the XGBoost-based multiparametric model significantly outperformed most single- and dual-sequence models in Centers 1 and 2. In Center 3, fewer differences reached statistical significance, indicating a narrower performance gap; nevertheless, the multiparametric approach retained high accuracy, highlighting robustness to inter-center variability. In addition, DCA demonstrated that the multiparametric fusion model yielded the highest net benefit across a wide range of threshold probabilities in all three centers, further supporting its potential clinical utility for individualized decision-making. The superior performance of XGBoost is attributable to its ability to model complex, non-linear feature interactions through gradient boosting decision trees. At the same time, built-in regularization mitigates overfitting in high-dimensional data (40-42). As radiomics often generates thousands of features-numerous redundant or highly associated-the sequential feature selection pipeline (variance thresholding, Pearson correlation filtering, univariate ANOVA and LASSO regression) effectively eliminated non-informative variables while retaining the most discriminative ones (43-45). This strategy yielded a final feature set dominated by clinical biomarkers and shape-related descriptors from DWI, consistent with prior evidence that tumor morphology and proliferation markers carry strong subtype-discriminative value (46-48). The proposed radiomics framework is designed for two complementary clinical scenarios. First, in pre-biopsy or biopsy-limited settings where Ki-67 and PR are unavailable, the pure imaging-based radiomics model (AUCs of 0.9088, 0.8621, and 0.8426 across three centers) can serve as a non-invasive tool for preliminary subtype stratification. Second, in post-biopsy settings where IHC biomarkers are available, the clinical-radiomics fusion model offers incremental value by capturing whole-tumor heterogeneity that localized biopsy sampling may miss(49). This is particularly relevant given that intratumoral heterogeneity can lead to discordance between biopsy and surgical specimens, potentially resulting in subtype misclassification based on IHC alone(50). This study has several limitations. First, the sample sizes of the external validation cohorts were relatively limited, particularly the luminal A subgroup in Center 2 (n=7), which may affect statistical power and subgroup-specific performance estimates. Although bootstrap resampling was performed to assess metric stability, future studies with larger and more balanced multicenter cohorts are needed to further validate the proposed model. Second, although standardized spatial resampling and z-score intensity normalization were applied to improve cross-center comparability, explicit harmonization methods such as ComBat were not employed due to the limited sample sizes in some external centers, and variations in scanner hardware and vendor-specific reconstruction algorithms may still influence radiomic feature stability. Third, although SMOTE was strictly applied only within training folds to prevent information leakage, synthetic oversampling in limited cohorts may still amplify noise. Fourth, formal inter- or intra-observer reproducibility analysis for tumor segmentation was not performed; future studies should incorporate multi-reader segmentation with intraclass correlation coefficient analysis. Fifth, as this was a retrospective study, the exact number of patients excluded at each step was not systematically recorded. Finally, the conventional radiomics and machine learning framework may overlook complex patterns that could be captured by deep learning approaches, and reliance on complete multi-sequence data limits applicability when sequences are missing. Future work should aim to address the limitations of the present study. First, although SMOTE was strictly applied only within training folds to prevent information leakage, synthetic oversampling in limited cohorts may still introduce noise. Expanding the dataset to include more centers, scanner types and acquisition protocols will strengthen feature robustness and improve generalizability. Incorporating scanner-specific harmonization techniques, such as ComBat or deep learning-based normalization, may further reduce inter-scanner variability. Second, the incorporation of radiomics-informed deep learning approaches and spatial/multi-scale features could capture more complex non-linear imaging patterns that conventional radiomics might miss, potentially complementing handcrafted features for enhanced tumor characterization. Finally, extending the model to predict additional clinically relevant endpoints, such as response to neoadjuvant chemotherapy, recurrence risk or metastasis probability, would broaden its clinical utility. These expansions, together with prospective multicenter validation, could facilitate the translation of the proposed model into real-world breast cancer management workflows. Declarations Ethics approval and consent to participate The present study was performed according to the Declaration of Helsinki and was approved by the institutional review boards of all participating institutions. Ethical approval was obtained from: (i) Guiyang Maternal and Child Health Care Hospital/Guiyang Children's Hospital (approval no. 2024-33; approved April 2024); (ii) Guiyang Second People's Hospital/The Affiliated Jinyang Hospital of Guizhou Medical University (approval no. 2022-2; approved March 2022); and (iii) Zunyi First People's Hospital/The Third Affiliated Hospital of Zunyi Medical University (approval no. 2024-1-439; approved August 2024). Due to the retrospective nature of the study and the use of anonymized patient data, the requirement for written informed consent was waived by all institutional review boards. All procedures complied with the Declaration of Helsinki and relevant national and institutional guidelines for research involving human subjects. Consent for publication Not applicable. Availability of data and materials The data generated in the present study may be requested from the corresponding author. Competing interests The authors declare that they have no competing interests. Funding The present work was supported by the Guizhou Provincial Basic Research Program (Natural Science; grant no. QKHJC-ZK-2023-006) and the Guiyang Health Bureau Science and Technology Project (grant no. ZWJKJ-2021-21). Authors' contributions HZ conceived the study and drafting of the manuscript. DY was responsible for data processing, including analysis and interpretation. XZ designed the study design, critically revised the manuscript. SL, XL and LW contributed to the acquisition of clinical and imaging data. HZ and DY confirm the authenticity of all the raw data. All authors read and approved the final version of the manuscript. Acknowledgements Not applicable. References International WCRF. Breast cancer statistics London: World Cancer Research Fund International; 2022 [Available from: https://www.wcrf.org/preventing-cancer/cancer-statistics/breast-cancer-statistics. Perou CM, Sørlie T, Eisen MB, Van De Rijn M, Jeffrey SS, Rees CA, et al. Molecular portraits of human breast tumours. nature. 2000;406(6797):747-52. Parker JS, Mullins M, Cheang MC, Leung S, Voduc D, Vickery T, et al. Supervised risk predictor of breast cancer based on intrinsic subtypes. Journal of clinical oncology. 2009;27(8):1160-7. Sørlie T, Perou CM, Tibshirani R, Aas T, Geisler S, Johnsen H, et al. Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proceedings of the National Academy of Sciences. 2001;98(19):10869-74. Giaquinto AN, Sung H, Miller KD, Kramer JL, Newman LA, Minihan A, et al. Breast cancer statistics, 2022. CA: a cancer journal for clinicians. 2022;72(6):524-41. Torre LA, Islami F, Siegel RL, Ward EM, Jemal A. Global cancer in women: burden and trends. Cancer epidemiology, biomarkers & prevention. 2017;26(4):444-57. Colleoni M, Sun Z, Price KN, Karlsson P, Forbes JF, Thürlimann B, et al. Annual hazard rates of recurrence for breast cancer during 24 years of follow-up: results from the international breast cancer study group trials I to V. Journal of clinical oncology. 2016;34(9):927-35. Bago-Horvath Z, Rudas M, Singer CF, Greil R, Balic M, Lax SF, et al. Predictive Value of Molecular Subtypes in Premenopausal Women with Hormone Receptor–positive Early Breast Cancer: Results from the ABCSG Trial 5. Clinical Cancer Research. 2020;26(21):5682-8. Zhang X, Yang Z, Cui W, Zheng C, Li H, Li Y, et al. Preoperative prediction of axillary sentinel lymph node burden with multiparametric MRI-based radiomics nomogram in early-stage breast cancer. European Radiology. 2021;31:5924-39. Zhang Y, Zhou Y, Mao F, Yao R, Sun Q. Ki‐67 index, progesterone receptor expression, histologic grade and tumor size in predicting breast cancer recurrence risk: A consecutive cohort study. Cancer Communications. 2020;40(4):181-93. Yuen S, Monzawa S, Yanai S, Matsumoto H, Yata Y, Ichinose Y, et al. The association between MRI findings and breast cancer subtypes: focused on the combination patterns on diffusion-weighted and T2-weighted images. Breast Cancer. 2020;27(5):1029-37. Montemurro F, Martincich L, Sarotto I, Bertotto I, Ponzone R, Cellini L, et al. Relationship between DCE-MRI morphological and functional features and histopathological characteristics of breast cancer. European radiology. 2007;17(6):1490-7. Duong DH, Nguyen QT, Nguyen DT, Tran TL, Duong TG. The association of magnetic resonance imaging features with five molecular subtypes of breast cancer. European Journal of Radiology Open. 2024;13:100585. Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, Van Stiphout RG, Granton P, et al. Radiomics: extracting more information from medical images using advanced feature analysis. European journal of cancer. 2012;48(4):441-6. Lambin P, Van Stiphout RG, Starmans MH, Rios-Velazquez E, Nalbantov G, Aerts HJ, et al. Predicting outcomes in radiation oncology—multifactorial decision support systems. Nature reviews Clinical oncology. 2013;10(1):27-40. Li C, Song L, Yin J. Intratumoral and peritumoral radiomics based on functional parametric maps from breast DCE‐MRI for prediction of HER‐2 and Ki‐67 status. Journal of Magnetic Resonance Imaging. 2021;54(3):703-14. Fan M, Cui Y, You C, Liu L, Gu Y, Peng W, et al. Radiogenomic signatures of oncotype DX recurrence score enable prediction of survival in estrogen receptor–positive breast cancer: A multicohort study. Radiology. 2022;302(3):516-24. Rizzo S, Botta F, Raimondi S, Origgi D, Fanciullo C, Morganti AG, et al. Radiomics: the facts and the challenges of image analysis. European radiology experimental. 2018;2:1-8. Xie T, Wang Z, Zhao Q, Bai Q, Zhou X, Gu Y, et al. Machine learning-based analysis of MR multiparametric radiomics for the subtype classification of breast cancer. Frontiers in oncology. 2019;9:505. Parmar C, Grossmann P, Bussink J, Lambin P, Aerts HJ. Machine learning methods for quantitative radiomic biomarkers. Scientific reports. 2015;5(1):1-11. Saha A, Harowicz MR, Grimm LJ, Kim CE, Ghate SV, Walsh R, et al. A machine learning approach to radiogenomics of breast cancer: a study of 922 subjects and 529 DCE-MRI features. British journal of cancer. 2018;119(4):508-16. Li H, Zhu Y, Burnside ES, Huang E, Drukker K, Hoadley KA, et al. Quantitative MRI radiomics in the prediction of molecular classifications of breast cancer subtypes in the TCGA/TCIA data set. NPJ breast cancer. 2016;2:16012. Ji Y, Whitney HM, Li H, Liu P, Giger ML, Zhang X. Differences in molecular subtype reference standards impact AI-based breast cancer classification with dynamic contrast-enhanced MRI. Radiology. 2023;307(1):e220984. Decoux A, Duron L, Habert P, Roblot V, Arsovic E, Chassagnon G, et al. Comparative performances of machine learning algorithms in radiomics and impacting factors. Scientific Reports. 2023;13(1):14069. Grimm LJ, Johnson KS, Marcom PK, Baker JA, Soo MS. Can breast cancer molecular subtype help to select patients for preoperative MR imaging? Radiology. 2015;274(2):352-8. LemaÃŽtre G, Nogueira F, Aridas CK. Imbalanced-learn: A python toolbox to tackle the curse of imbalanced datasets in machine learning. Journal of machine learning research. 2017;18(17):1-5. Hosmer Jr DW, Lemeshow S, Sturdivant RX. Applied logistic regression: John Wiley & Sons; 2013. Peterson LE. K-nearest neighbor. Scholarpedia. 2009;4(2):1883. Webb GI. Naïve bayes. Encyclopedia of machine learning and data mining: Springer; 2017. p. 895-6. Hearst MA, Dumais ST, Osuna E, Platt J, Scholkopf B. Support vector machines. IEEE Intelligent Systems and their applications. 1998;13(4):18-28. Song Y-Y, Lu Y. Decision tree methods: applications for classification and prediction. Shanghai archives of psychiatry. 2015. Breiman L. Random forests. Machine learning. 2001;45(1):5-32. Margineantu DD, Dietterich TG, editors. Pruning adaptive boosting. ICML; 1997. Friedman JH. Greedy function approximation: a gradient boosting machine. Annals of statistics. 2001:1189-232. Chen T, Guestrin C, editors. Xgboost: A scalable tree boosting system. Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining; 2016. Han B, Pan S, Ruan M. Predictive value of neoadjuvant chemotherapy on breast cancer subtypes evaluated by DCE-MRI and MRI perfusion parameters. Indian Journal of Experimental Biology (IJEB). 2023;61(10). Luo H-B, Du M-Y, Liu Y-Y, Wang M, Qing H-M, Wen Z-p, et al. Differentiation between luminal a and B molecular subtypes of breast cancer using pharmacokinetic quantitative parameters with histogram and texture features on preoperative dynamic contrast-enhanced magnetic resonance imaging. Academic Radiology. 2020;27(3):e35-e44. Montemezzi S, Camera L, Giri MG, Pozzetto A, Caliò A, Meliadò G, et al. Is there a correlation between 3T multiparametric MRI and molecular subtypes of breast cancer? European Journal of Radiology. 2018;108:120-7. Huang Y, Wang X, Cao Y, Li M, Li L, Chen H, et al. Multiparametric MRI model to predict molecular subtypes of breast cancer using Shapley additive explanations interpretability analysis. Diagnostic and interventional imaging. 2024;105(5):191-205. Zhou Y, Lin G, Chen W, Chen Y, Shi C, Peng Z, et al. Multiparametric MRI-based Interpretable Machine Learning Radiomics Model for Distinguishing Between Luminal and Non-luminal Tumors in Breast Cancer: A Multicenter Study. Academic Radiology. 2025. Vamvakas A, Tsivaka D, Logothetis A, Vassiou K, Tsougos I. Breast cancer classification on multiparametric MRI–increased performance of boosting ensemble methods. Technology in cancer research & treatment. 2022;21:15330338221087828. Yimit Y, Yasin P, Tuersun A, Wang J, Wang X, Huang C, et al. Multiparametric MRI-based interpretable radiomics machine learning model differentiates medulloblastoma and ependymoma in children: a two-center study. Academic Radiology. 2024;31(8):3384-96. Demircioğlu A. Benchmarking feature selection methods in radiomics. Investigative radiology. 2022;57(7):433-43. Zhang W, Guo Y, Jin Q. Radiomics and its feature selection: a review. Symmetry. 2023;15(10):1834. Perniciano A, Loddo A, Di Ruberto C, Pes B. Insights into radiomics: impact of feature selection and classification. Multimedia Tools and Applications. 2024:1-27. Huang T, Fan B, Qiu Y, Zhang R, Wang X, Wang C, et al. Application of DCE-MRI radiomics signature analysis in differentiating molecular subtypes of luminal and non-luminal breast cancer. Frontiers in Medicine. 2023;10:1140514. Zhang S, Wang X, Yang Z, Zhu Y, Zhao N, Li Y, et al. Intra-and peritumoral radiomics model based on early DCE-MRI for preoperative prediction of molecular subtypes in invasive ductal breast carcinoma: a multitask machine learning study. Frontiers in Oncology. 2022;12:905551. Leithner D, Mayerhoefer ME, Martinez DF, Jochelson MS, Morris EA, Thakur SB, et al. Non-invasive assessment of breast cancer molecular subtypes with multiparametric magnetic resonance imaging radiomics. Journal of clinical medicine. 2020;9(6):1853. Ye D-M, Wang H-T, Yu T. The application of radiomics in breast MRI: a review. Technology in Cancer Research & Treatment. 2020;19:1533033820916191. Ahn S, Lee J, Cho M-S, Park S, Sung SH. Evaluation of Ki-67 index in core needle biopsies and matched breast cancer surgical specimens. Archives of pathology & laboratory medicine. 2018;142(3):364-8. Tables Table 1. Baseline clinicopathological characteristics of patients in Centers 1, 2 and 3. Characteristic Center 1 Center 2 Center 3 Luminal A (n=19) Luminal B (n=45) P-value Luminal A (n=7) Luminal B (n=29) P-value Luminal A (n=18) Luminal B (n=12) P-value Sex - - 0.654 Female 19 (100%) 45 (100%) - 7 (100%) 29 (100%) - 16(88.89%) 12(100%) - Male 0(0.00%) 0(0.00%) - 0(0.00%) 0(0.00%) - 2(11.11%) 0(0.00%) - Age, years 57.05±10.50 49.00±9.99 0.0076 55.00±13.52 51.34±12.93 0.5337 51.33±12.70 54.25±12.30 0.536 Histological type - 0.4802 0.640 1 19 (100%) 45 (100%) - 6(85.71%) 27(93.10%) - 13(72.22%) 10(83.33%) - 2 0 (0.00%) 0(0.00%) - 0(0.00%) 1(3.45%) - 1(5.56%) 0(0.00%) - 3 0(0.00%) 0(0.00%) - 1(14.29%) 1(3.45%) - 4(22.22%) 2(16.67%) - Tumor margins 0.6540 1.0 >0.9999 0 12(63.16%) 24(53.33%) - 4(57.14%) 16(55.17%) - 8(44.44%) 6(50.00%) - 1 7(36.84%) 21(46.67%) - 3(42.86%) 13(44.83%) - 10(55.56%) 6(50.00%) - Tumor morphology 0.3961 0.1625 >0.9999 0 4(21.05%) 16(35.56%) - 2(28.57%) 1(3.45%) - 5(27.78%) 4(33.33%) - 1 15(78.95%) 29(64.44%) - 5(71.43%) 28(96.55%) - 13(72.22%) 8(66.67%) - ER, %) 79.73±18.74 67.00±22.82 0.0252 81.43±15.74 78.45±22.32 0.6879 84.44±13.27 51.67±40.47 0.018 PR, % 66.05±18.90 27.07±27.59 <0.0001 43.57±36.82 47.07±38.63 0.8279 76.67±12.49 28.75±30.07 0.001 HER-2 0.2006 0.7429 0.027 0 7(36.84%) 11(24.44%) - 3(42.86%) 10(34.48%) - 9(50.00%) 5(41.67%) - 1 8(42.11%) 20(44.44%) - 4(57.14%) 17(58.62%) - 9(50.00%) 3(25.00%) - 2 4(21.05%) 6(13.33%) - 0(0.00%) 2(6.90%) - 0(0.00%) 0(0.00%) - 3 0(0.00%) 8(17.78%) - 0(0.00%) 0(0.00%) - 0(0.00%) 4(33.33%) - Ki-67, %) 16.58±6.02 34.31±20.60 <0.0001 6.86±3.67 35.00±20.04 <0.0001 8.89±4.71 12.92±10.33 0.226 Lymph node status 0.1795 0.4968 0.594 0 10 (52.63%) 14(31.11%) - 2(28.57%) 15(51.72%) - 12(66.67%) 6(50.00%) - 1 9(47.37%) 31(68.89%) - 5(71.43%) 14(48.28%) - 6(33.33%) 6(50.00%) - Baseline clinicopathological characteristics of patients in the training and test cohorts. Data are presented as mean ± standard deviation or as n (%). – indicates not applicable. P - values were calculated using the independent-samples t-test for continuous variables and the χ 2 test or Fisher’s exact test for categorical variables, as appropriate. Table 2. Diagnostic performance of different MRI modalities in differentiating luminal A from luminal B breast cancers in Center 1. Modality Sensitivity Specificity Accuracy PPV NPV AUC DCE 0.6889 (0.5500, 0.8222) 0.8947 (0.7333, 1.0000) 0.7500 (0.6406, 0.8438) 0.9394 (0.8462, 1.0000) 0.5484 (0.3824, 0.7273) 0.8082 (0.6910, 0.9103) T2 0.6889 (0.5555, 0.8182) 0.8421 (0.6663, 1.0000) 0.7344 (0.6250, 0.8285) 0.9118 (0.8107, 1.0000) 0.5333 (0.3571, 0.7200) 0.8018 (0.6860, 0.9015) DWI 0.7111 (0.5743, 0.8409) 0.8947 (0.7391, 1.0000) 0.7656 (0.6719, 0.8594) 0.9412 (0.8571, 1.0000) 0.5667 (0.3714, 0.7500) 0.8363 (0.7168, 0.9341) T2-DCE 0.5778 (0.4389, 0.7209) 0.7895 (0.5833, 0.9524) 0.6406 (0.5156, 0.7500) 0.8667 (0.7308, 0.9677) 0.4412 (0.2666, 0.5947) 0.7427 (0.6172, 0.8557) T2-DWI 0.7556 (0.6250, 0.8810) 0.8947 (0.7368, 1.0000) 0.7969 (0.7031, 0.8906) 0.9444 (0.8610, 1.0000) 0.6071 (0.4090, 0.7778) 0.8772 (0.7765, 0.9527) DCE-DWI 0.7556 (0.6249, 0.8751) 0.9474 (0.8235, 1.0000) 0.8125 (0.7188, 0.9062) 0.9714 (0.9062, 1.0000) 0.6207 (0.4375, 0.7826) 0.8655 (0.7578, 0.9512) T2-DCE-DWI 0.7778 (0.6591, 0.8914) 0.9474 (0.8235, 1.0000) 0.8281 (0.7344, 0.9219) 0.9722 (0.9024, 1.0000) 0.6429 (0.4687, 0.8078) 0.9088 (0.8313, 0.9758) T2-DCE-DWI-clinical 0.8000 (0.6807, 0.9000) 0.9474 (0.8180, 1.0000) 0.8438 (0.7500, 0.9219) 0.9730 (0.9143, 1.0000) 0.6667 (0.4848, 0.8400) 0.9392 (0.8682, 0.9859) Values are presented as mean (95% confidence interval). Sensitivity, specificity, PPV and NPV were calculated using the luminal B subtype as the positive class. DCE-MRI; dynamic contrast-enhanced magnetic resonance imaging; T2WI, T2-weighted imaging; DWI, diffusion-weighted imaging; PPV, positive predictive value; NPV, negative predicative value; AUC, area under the curve. Table 3. Diagnostic performance of different MRI modalities in Center 2. Modality Sensitivity Specificity Accuracy PPV NPV AUC DCE 0.8621 (0.7306, 0.9677) 0.5714 (0.1667, 1.0000) 0.8056 (0.6667, 0.9167) 0.8929 (0.7600, 1.0000) 0.5000 (0.1667, 0.8580) 0.7685 (0.5444, 0.9697) T2 0.8621 (0.7272, 0.9677) 0.5714 (0.1667, 1.0000) 0.8056 (0.6667, 0.9167) 0.8929 (0.7692, 1.0000) 0.5000 (0.1429, 0.8571) 0.7340 (0.4285, 0.9609) DWI 0.6897 (0.5333, 0.8519) 0.7143 (0.3333, 1.0000) 0.6944 (0.5556, 0.8333) 0.9091 (0.7777, 1.0000) 0.3571 (0.1000, 0.6250) 0.7463 (0.4318, 1.0000) T2-DCE 0.7931 (0.6428, 0.9286) 0.5714 (0.1533, 1.0000) 0.7500 (0.6111, 0.8889) 0.8846 (0.7600, 1.0000) 0.4000 (0.1111, 0.7143) 0.7734 (0.4683, 1.0000) T2-DWI 0.7931 (0.6296, 0.9286) 0.8571 (0.5000, 1.0000) 0.8056 (0.6667, 0.9167) 0.9583 (0.8571, 1.0000) 0.5000 (0.2000, 0.7857) 0.8128 (0.4853, 0.9922) DCE-DWI 0.7931 (0.6333, 0.9286) 0.5714 (0.1667, 1.0000) 0.7500 (0.6111, 0.8889) 0.8846 (0.7600, 1.0000) 0.4000 (0.1000, 0.7276) 0.8399 (0.6617, 0.9798) T2-DCE-DWI 0.8621 (0.7143, 0.9677) 0.8571 (0.5000, 1.0000) 0.8611 (0.7500, 0.9722) 0.9615 (0.8750, 1.0000) 0.6000 (0.2727, 0.8889) 0.8621 (0.5911, 1.0000) T2-DCE-DWI-clinical 0.9310 (0.8276, 1.0000) 0.8571 (0.5000, 1.0000) 0.9167 (0.8056, 1.0000) 0.9643 (0.8846, 1.0000) 0.7500 (0.4000, 1.0000) 0.8966 (0.7059, 1.0000) Values are presented as mean (95% confidence interval). Sensitivity, specificity, PPV and NPV were calculated using the luminal B subtype considered as the positive class. DCE-MRI; dynamic contrast-enhanced magnetic resonance imaging; T2WI, T2-weighted imaging; DWI, diffusion-weighted imaging; PPV, positive predictive value; NPV, negative predicative value; AUC, area under the curve. Table 4. Diagnostic performance of different MRI modalities in Center 3. Modality Sensitivity Specificity Accuracy PPV NPV AUC DCE 0.6667 (0.3750, 0.9286) 0.5556 (0.3330, 0.7895) 0.6000 (0.4333, 0.7667) 0.5000 (0.2628, 0.7500) 0.7143 (0.4615, 0.9286) 0.7685 (0.5700, 0.9277) T2 0.6667 (0.3636, 0.9167) 0.8333 (0.6471, 1.0000) 0.7667 (0.6000, 0.9000) 0.7273 (0.4000, 1.0000) 0.7895 (0.5882, 0.9474) 0.7083 (0.4976, 0.8929) DWI 0.6667 (0.3846, 0.9093) 0.8333 (0.6316, 1.0000) 0.7667 (0.6000, 0.9000) 0.7273 (0.4286, 1.0000) 0.7895 (0.5556, 0.9474) 0.7593 (0.5486, 0.9398) T2-DCE 0.6667 (0.3844, 0.9167) 0.7778 (0.5714, 0.9444) 0.7333 (0.5667, 0.9000) 0.6667 (0.3636, 0.9231) 0.7778 (0.5882, 0.9474) 0.8009 (0.6294, 0.9500) T2-DWI 0.5833 (0.2857, 0.8571) 0.7222 (0.5000, 0.9286) 0.6667 (0.5000, 0.8333) 0.5833 (0.3000, 0.8750) 0.7222 (0.5000, 0.9286) 0.7593 (0.5503, 0.9260) DCE-DWI 0.7500 (0.4545, 1.0000) 0.6667 (0.4284, 0.8824) 0.7000 (0.5333, 0.8667) 0.6000 (0.3635, 0.8462) 0.8000 (0.5554, 1.0000) 0.7315 (0.5177, 0.9051) T2-DCE-DWI 0.8333 (0.6000, 1.0000) 0.7222 (0.5263, 0.9167) 0.7667 (0.6333, 0.9000) 0.6667 (0.4286, 0.9091) 0.8667 (0.6667, 1.0000) 0.8426 (0.6832, 0.9778) T2-DCE-DWI-clinical 0.8333 (0.5714, 1.0000) 0.7222 (0.5000, 0.9231) 0.7667 (0.6325, 0.9000) 0.6667 (0.4209, 0.9091) 0.8667 (0.6667, 1.0000) 0.8704 (0.7083, 0.9861) Values are presented as mean (95% confidence interval). Sensitivity, specificity, PPV and NPV were calculated using the luminal B subtype considered as the positive class. DCE-MRI; dynamic contrast-enhanced magnetic resonance imaging; T2WI, T2-weighted imaging; DWI, diffusion-weighted imaging; Clinical: clinical features; PPV, positive predictive value; NPV, negative predicative value; AUC, area under the curve. Additional Declarations No competing interests reported. Supplementary Files SupplementFig.1.png Supplement Fig.1 Heatmaps of pairwise DeLong's test P-values for AUC comparisons among machine learning classifiers. (a) Center 1, (b) Center 2, (c) Center 3. Color intensity represents the P-value, with darker colors indicating more significant differences. Values shown are P-values from DeLong's test. P < 0.05 was considered statistically significant. supplementFigure2.jpg Supplement Fig. 2 SHAP (SHapley Additive exPlanations) summary plots for the clinical-radiomics fusion model across three centers. (a)Center 2, (b) Center 3. Features are ranked by mean absolute SHAP value in descending order. Consistent high-ranking features across cohorts (Ki-67, PR status, age, and DWI_original_shape_MeshVolume) indicate stable and generalizable predictors. Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 08 May, 2026 Reviews received at journal 30 Apr, 2026 Reviewers agreed at journal 24 Apr, 2026 Reviews received at journal 22 Apr, 2026 Reviews received at journal 20 Apr, 2026 Reviewers agreed at journal 15 Apr, 2026 Reviewers agreed at journal 15 Apr, 2026 Reviewers invited by journal 15 Apr, 2026 Editor invited by journal 23 Mar, 2026 Editor assigned by journal 06 Feb, 2026 Submission checks completed at journal 06 Feb, 2026 First submitted to journal 05 Feb, 2026 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-8793807","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":627941612,"identity":"1bc5e40b-b1f7-490b-b36c-905f3f8b3ace","order_by":0,"name":"Hui zhou","email":"","orcid":"","institution":"The Affiliated Jinyang Hospital of Guizhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Hui","middleName":"","lastName":"zhou","suffix":""},{"id":627941613,"identity":"c562c6ea-3975-4930-8049-342267067ff9","order_by":1,"name":"Daoyu Yang","email":"","orcid":"","institution":"Guizhou University","correspondingAuthor":false,"prefix":"","firstName":"Daoyu","middleName":"","lastName":"Yang","suffix":""},{"id":627941614,"identity":"cf3384ca-b5c2-4db9-b679-9dca18e7bd83","order_by":2,"name":"Shiguang Li","email":"","orcid":"","institution":"The Affiliated Jinyang Hospital of Guizhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Shiguang","middleName":"","lastName":"Li","suffix":""},{"id":627941615,"identity":"58e4e9d5-b40c-4c34-91bf-4c723736520d","order_by":3,"name":"Xudong Liu","email":"","orcid":"","institution":"The Affiliated Jinyang Hospital of Guizhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xudong","middleName":"","lastName":"Liu","suffix":""},{"id":627941616,"identity":"e798ba2e-5e9f-4c2d-9b4c-3c0240192d88","order_by":4,"name":"Ling Wei","email":"","orcid":"","institution":"Zunyi First People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Ling","middleName":"","lastName":"Wei","suffix":""},{"id":627941617,"identity":"3ef2ca65-bccc-40a5-a160-8df32ec797c1","order_by":5,"name":"Jiarui Wang","email":"","orcid":"","institution":"Guiyang Maternal and Child Health Care Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jiarui","middleName":"","lastName":"Wang","suffix":""},{"id":627941618,"identity":"6e2675c0-78d2-43d6-a2e5-c99f989aaba3","order_by":6,"name":"Xianchun Zeng","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABDklEQVRIiWNgGAWjYNACAyCWAOIEHht+sEBCAbFaPsikSTaAtRgQYxNQC+MMm8MQLQx4tMi39x5+zVNgl8c/u8fsMU/OeQl+iezEDw8MGOT5xQ5g1cLYcy7NcoZBcrHEnTPmxjxnbktI9pzdLAF0mOHM2QlYtTBL5JgZfDBgTtwAZEjz9tyuMzjeuwGkJcHgNnYtbPJvzAwSDOqhWv6dkzA4zLv5Bz4tPBI8xg8+GBwGa5GcwXNAAmjLNry2SPDkmDHOMDieOONGWpnEB55kkF+2WSQYSOD0i3z7GePPPH+qE/tnJAMN57EDhlju5ps/Kmzk+aWxawF5RwKr9biUgwDzB3yyo2AUjIJRMAoYAPoWWLG0sOEtAAAAAElFTkSuQmCC","orcid":"","institution":"Guizhou Provincial People's Hospital","correspondingAuthor":true,"prefix":"","firstName":"Xianchun","middleName":"","lastName":"Zeng","suffix":""}],"badges":[],"createdAt":"2026-02-05 07:54:25","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8793807/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8793807/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107675583,"identity":"3c90ae93-c82b-464f-ac9d-e7a92d564406","added_by":"auto","created_at":"2026-04-24 00:44:17","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":245538,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of patient inclusion and dataset allocation. A total of 136 patients with preoperative dynamic contrast-enhanced-MRI, diffusion-weighted imaging and T2-weighted imaging, and pathologically confirmed luminal A or luminal B invasive ductal carcinoma were initially enrolled. After exclusion of 6 patients due to incomplete or missing pathology data, 130 patients were included in the final cohort. The cohort was divided into Center 1 (n=64; luminal A, n=19; luminal B, n=45), Center 2 (n=36; luminal A, n=7; luminal B, n=29) and Center 3 (n=30; luminal A, n=18; luminal B, n=12). MRI, magnetic resonance imaging.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8793807/v1/d2d428da5209f7f92c03a8ed.png"},{"id":107675585,"identity":"354d726c-097b-46a9-b8d2-a297fab01526","added_by":"auto","created_at":"2026-04-24 00:44:17","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1347313,"visible":true,"origin":"","legend":"\u003cp\u003eROI delineation on multiparametric breast MRI. (A) T2WI, (B) DWI and (C) DCE-MRI images from the same patient with pathologically confirmed invasive ductal carcinoma (luminal A subtype). The tumor is manually segmented (red mask) on each sequence to define the ROI for subsequent radiomics feature extraction. ROI, region of interest; DCE-MRI, dynamic contrast-enhanced magnetic resonance imaging; T2WI, T2-weighted imaging; DWI, diffusion-weighted imaging.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8793807/v1/58d9612f84813be68e1912ba.png"},{"id":107708030,"identity":"d7a2101d-ee06-49cc-936d-f04c6f1eeba2","added_by":"auto","created_at":"2026-04-24 09:21:42","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":4170563,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP summary plot illustrating the contribution of each feature to the model’s output. Features are ranked by their mean absolute SHAP values, with higher-ranked features exerting greater influence on model prediction. Each dot represents a single patient, where the horizontal position indicates the SHAP value (impact on model output) and the color denotes the feature value (red, high; blue, low). SHAP, SHapley Additive exPlanations; DCE, dynamic contrast-enhanced; PR, progesterone receptor; ER, estrogen receptor; HER2, human epidermal growth factor receptor 2.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-8793807/v1/d9ea912bc6a4dce2d95e7322.png"},{"id":107675587,"identity":"f78243ce-bda6-4478-af71-fcd05c19eb54","added_by":"auto","created_at":"2026-04-24 00:44:17","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":14799692,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves of machine learning classifiers across the three centers.(a) Center 1, (b) Center 2 and (c) Center 3. ROC curves are shown for nine classifiers: Logistic regression, KNN, naive Bayes, SVM, decision tree, random forest, AdaBoost, GBDT and XGBoost. AUC values are provided in the legend for each model. ROC, receiver operating characteristic; AUC, area under the curve; KNN, k-nearest neighbors; SVM, support vector machine; AdaBoost, Adaptive Boosting; GBDT, gradient boosting decision tree; XGBoost, Extreme Gradient Boosting.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8793807/v1/bbc3194ce6fd2a7276e5d61a.png"},{"id":107707845,"identity":"a1e487ba-4008-4425-96c6-a450e0528bdd","added_by":"auto","created_at":"2026-04-24 09:21:16","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":9773009,"visible":true,"origin":"","legend":"\u003cp\u003eReceiver operating characteristic curves of the Extreme Gradient Boosting model for differentiating luminal A from luminal B breast cancers across different centers. (a) Center 1, (b) Center 2 and (c) Center 3. AUC, area under the curve. Model 1, DCE based radiomics; Model 2 (T2WI-based radiomics); Model 3 (DWI-based radiomics); Model 4 (T2WI+DCE fusion); Model 5 (T2WI+DWI fusion); Model 6 (DCE+DWI fusion); Model 7 (T2WI+DCE+DWI fusion); Model 8 (T2WI+DCE+DWI+Clinical fusion) DCE, dynamic contrast-enhanced magnetic resonance imaging; T2WI, T2-weighted imaging; DWI, diffusion-weighted imaging.\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-8793807/v1/d9d86315063bd965b420d152.png"},{"id":107675589,"identity":"0d0a3513-b8b3-4c8c-88ff-76e95ad947f9","added_by":"auto","created_at":"2026-04-24 00:44:17","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1541620,"visible":true,"origin":"","legend":"\u003cp\u003eDecision curve analysis of the Extreme Gradient Boostingmodel for differentiating luminal A from luminal B breast cancers across different centers. (a) Center 1, (b) Center 2 and (c) Center 3. Model 1 (DCE based radiomics); Model 2 (T2WI-based radiomics); Model 3 (DWI-based radiomics); Model 4 (T2WI+DCE fusion); Model 5 (T2WI+DWI fusion); Model 6 (DCE+DWI fusion); Model 7 (T2WI+DCE+DWI fusion); Model 8 (T2WI+DCE+DWI+Clinical fusion).\u003c/p\u003e","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-8793807/v1/e2257fce84d01b37ce858020.png"},{"id":107675590,"identity":"977134d1-a4aa-4490-8545-b69aa19f3653","added_by":"auto","created_at":"2026-04-24 00:44:17","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":1524060,"visible":true,"origin":"","legend":"\u003cp\u003ePairwise comparisons of model performance using DeLong tests across Centers 1-3. Pairwise comparisons of diagnostic performance among eight model configurations using the Extreme Gradient Boostingclassifier across (a) Center 1 (b) Center 2 and (c) Center 3. Each cell represents the P-value from the DeLong test comparing the AUCs of two models, with warmer colors indicating higher P-values (less significant difference) and cooler colors indicating lower P-values (more substantial difference). In Centers 1 and 2, the multiparametric clinical-radiomics fusion model significantly outperformed most single- and dual-sequence models (P\u0026lt;0.05 in the majority of comparisons). By contrast, in Center 3, fewer pairwise differences reached statistical significance, indicating smaller performance gaps among models. Model 1 (DCE based radiomics); Model 2 (T2WI-based radiomics); Model 3 (DWI-based radiomics); Model 4 (T2WI+DCE fusion); Model 5 (T2WI+DWI fusion); Model 6 (DCE+DWI fusion); Model 7 (T2WI+DCE+DWI fusion); Model 8 (T2WI+DCE+DWI+Clinical fusion).\u003c/p\u003e","description":"","filename":"Figure7.png","url":"https://assets-eu.researchsquare.com/files/rs-8793807/v1/62a78f098d551c6b3bd9d38a.png"},{"id":107709526,"identity":"ac8c784a-a311-4761-aaee-44b4a1d02ded","added_by":"auto","created_at":"2026-04-24 09:36:14","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":29034341,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8793807/v1/c440a15f-283d-4d10-ae26-5029d0ed776d.pdf"},{"id":107708106,"identity":"391c4270-dc51-426b-866b-ac4b9ab40b20","added_by":"auto","created_at":"2026-04-24 09:21:56","extension":"png","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1005965,"visible":true,"origin":"","legend":"\u003cp\u003eSupplement Fig.1 Heatmaps of pairwise DeLong's test P-values for AUC comparisons among machine learning classifiers. (a) Center 1, (b) Center 2, (c) Center 3. Color intensity represents the P-value, with darker colors indicating more significant differences. Values shown are P-values from DeLong's test. P \u0026lt; 0.05 was considered statistically significant.\u003c/p\u003e","description":"","filename":"SupplementFig.1.png","url":"https://assets-eu.researchsquare.com/files/rs-8793807/v1/8b98a427fcd55b7a60fac036.png"},{"id":107707754,"identity":"f6c23027-f5a2-4bb1-b5b9-95fcaf8c3978","added_by":"auto","created_at":"2026-04-24 09:21:04","extension":"jpg","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":684845,"visible":true,"origin":"","legend":"\u003cp\u003eSupplement Fig. 2 SHAP (SHapley Additive exPlanations) summary plots for the clinical-radiomics fusion model across three centers. (a)Center 2, (b) Center 3. Features are ranked by mean absolute SHAP value in descending order. Consistent high-ranking features across cohorts (Ki-67, PR status, age, and DWI_original_shape_MeshVolume) indicate stable and generalizable predictors.\u003c/p\u003e","description":"","filename":"supplementFigure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8793807/v1/52dd0a060801b8a70de08704.jpg"}],"financialInterests":"No competing interests reported.","formattedTitle":"Multiparametric MRI-based radiomics combined with machine learning for preoperative differentiation of luminal A and luminal B breast cancer: a multicenter study","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAccording to 2022 global cancer statistics, ~\u0026thinsp;2.3\u0026nbsp;million women worldwide were diagnosed with breast cancer, underscoring its status as a major public health concern(\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). The Prediction Analysis of Microarray 50 (PAM50) gene expression assay categorizes breast cancer tumors into five intrinsic subtypes-luminal A, luminal B, human epidermal growth factor receptor 2 (HER2)-enriched, basal-like and normal-like-based on the messenger ribonucleic acid (mRNA) expression of 50 genes (\u003cspan additionalcitationids=\"CR3\" citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e). Among these, luminal A and B are the most common subtypes (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e), but they differ markedly in treatment approaches and prognosis. Despite sharing hormone receptor positivity, luminal B tumors exhibit higher proliferative activity and lower hormone receptor expression, resulting in worse prognosis and higher recurrence rates compared with luminal A (\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e). Accurate preoperative differentiation between these two subtypes is crucial for tailoring treatment and improving patient outcomes. However, current classification methods, although effective, often rely on invasive, costly, and time-consuming procedures, such as histopathological biopsy and molecular profiling (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eMagnetic resonance imaging (MRI) techniques such as T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI) and dynamic contrast-enhanced MRI (DCE-MRI) have emerged as promising non-invasive tools for breast tumor assessment. These modalities offer complementary information, with T2WI reflecting anatomical structures and tissue contrast, DWI providing insights into cellular density and microstructural integrity and DCE-MRI capturing perfusion characteristics and vascular dynamics (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e). Nevertheless, their standalone diagnostic performance remains limited, particularly in capturing tumor heterogeneity and reliably distinguishing between luminal A and luminal B subtypes (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). This limitation underscores the need for advanced analytical approaches to integrate multiparametric MRI data to enhance diagnostic precision and facilitate personalized treatment decision-making.\u003c/p\u003e \u003cp\u003eRadiomics, introduced by Lambin(\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e), has emerged as a valuable tool for extracting high-throughput quantitative features from medical images, converting them into data that can be analyzed for tumor heterogeneity (\u003cspan additionalcitationids=\"CR16\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e). By converting medical images into structured data, radiomics facilitates objective tumor characterization beyond visual interpretation. When combined with machine learning algorithms such as support vector machines and random forests, radiomics has shown potential in tumor classification and outcome prediction (\u003cspan additionalcitationids=\"CR19 CR20\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e). Despite its success in various oncologic contexts, studies specifically addressing the classification of luminal A and B breast cancer subtypes remain limited. These two subtypes exhibit overlapping imaging phenotypes but differ notably in molecular expression profiles, treatment response and prognostic outcomes, posing a considerable diagnostic challenge. Accurate non-invasive differentiation between these subtypes is crucial for guiding personalized treatment decisions, particularly when biopsy is contraindicated or with insufficient tissue sampling.\u003c/p\u003e \u003cp\u003eAlthough radiomics-based approaches have been explored for breast cancer molecular subtyping, several gaps remain. First, most studies have focused on distinguishing triple-negative or HER2-enriched subtypes from others, few studies have specifically addressed the clinically challenging task of differentiating luminal A from luminal B subtypes (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e). Second, the majority of existing studies were conducted at single centers using single MRI sequences, limiting the generalizability of the findings(\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e). Third, systematic comparisons of multiple machine learning classifiers using identical feature sets are rarely reported, making it difficult to identify the optimal modeling strategy(\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e). Fourth, although explainability techniques such as SHAP have been applied in breast cancer imaging, their integration with multiparametric MRI radiomics and multicenter validation for luminal subtype classification has not been explored(\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTherefore, the present study aimed to develop and validate a radiomics-based multiparametric model that integrates MRI-derived imaging features, clinical characteristics, pathological markers [estrogen receptor (ER), progesterone receptor (PR), HER2, Ki-67] and lymph node status to distinguish luminal A from luminal B breast cancers. By leveraging multi-center datasets and adopting an integrated radio-genomic approach, the present study aimed to enhance the robustness, accuracy and generalizability of preoperative subtype classification to facilitate personalized treatment planning.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cp\u003e\u003cem\u003eData acquisition\u003c/em\u003e. This retrospective multicenter study was conducted at three institutions (Center 1, Center 2, and Center 3). A total of 136 anonymized breast cancer cases diagnosed between October 2020 and June 2024 were included. \u003c/p\u003e\n\u003cp\u003eAll patients underwent preoperative MRI, including T2WI, DWI and DCE-MRI. The detailed scanning protocols are provided in Supplement Table S1. After excluding patients with incomplete imaging data, 130 patients with pathologically confirmed luminal A or luminal B invasive ductal carcinoma and complete multimodal MRI were included in the final analysis. The inclusion criteria were as follows: (i) Age 18-80 years; (ii) histopathologically confirmed luminal A or B invasive ductal carcinoma; (iii) complete preoperative breast MRI including T2WI, DWI and DCE-MRI; (iv) available clinicopathological data (ER, PR, HER2, Ki-67); and (v) treatment-naive status. The exclusion criteria were as follows: (i) Patients with incomplete imaging data across any of the three MRI modalities; (ii) poor image quality affecting feature extraction; (iii) previous breast cancer treatment; and (iv) missing critical pathological markers.\u003c/p\u003e\n\u003cp\u003eSubsequently, data from Center 1, comprising 64 patients, were used for model development and internal evaluation using five-fold cross-validation. Data from the remaining two institutions, Center 2 (n=36) and Center 3 (n=30), served as independent external validation cohorts. Fig. 1 illustrates the flowchart detailing the inclusion and exclusion process for the study cohort. \u003c/p\u003e\n\n\u003cp\u003e\u003cem\u003eImage segmentation and radiomics feature extraction.\u003c/em\u003e In the present study, tumor segmentation was performed manually on a slice-by-slice basis (2D approach) across three imaging modalities (DCE-MRI, T2WI, DWI) using ITK-SNAP software (version 3.8.0; itksnap.org) (Fig. 2). For each case, the tumor was delineated on all axial slices where the lesion was visible, and the individual 2D contours were subsequently combined to generate a three-dimensional volume of interest (ROI) for radiomics feature extraction. All delineations were subsequently reviewed and verified by a senior radiologist with over 20 years of experience to ensure anatomical accuracy and consistency.Radiomics features were extracted from these ROIs using the PyRadiomics package (version 3.0; pyradiomics.readthedocs.io), covering a wide range of descriptors: First-order statistics reflecting intensity distribution; shape-based features characterizing tumor geometry; and texture features derived from gray-level co-occurrence matrix, gray-level run length matrix, gray-level size zone matrix (GLSZM), gray-level dependence matrix and neighboring gray-tone difference matrix. Prior to feature extraction, all MRI images were resampled to an isotropic voxel spacing of 1 \u0026times; 1 \u0026times; 1 mm\u0026sup3; using B-spline interpolation to minimize the influence of voxel size heterogeneity across centers. Voxel intensities were subsequently standardized using z-score normalization within the tumor volume (mean = 0, standard deviation = 1). Gray-level discretization was then performed using a fixed bin width of 25 to ensure consistent texture quantization across all cases. To improve tumor heterogeneity and spatial complexity sensitivity, image filters including Laplacian of Gaussian filters with s values of 1.0, 2.0 and 3.0 and multi-level wavelet decompositions were applied before feature extraction. All extracted radiomics features were further standardized using z-score scaling before model construction to improve feature comparability across centers.\u003c/p\u003e\n\n\u003cp\u003e\u003cem\u003eFeatures selection and model construction\u003c/em\u003e. A multi-step feature selection procedure was employed to optimize model input, with all steps performed strictly within the training folds of five-fold cross-validation. Initially, features exhibiting low variance were excluded using a variance threshold method (threshold, 0.01) to remove uninformative variables. Subsequently, highly correlated features were identified using pairwise Pearson correlation analysis. For each feature pair with a correlation coefficient greater than 0.9, the feature with the lower univariate analysis of variance (ANOVA) F-value was removed while the more informative feature was retained, thereby mitigating redundancy and multicollinearity. The remaining feature set was refined through univariate analysis employing the ANOVA F-test, selecting the top features (up to 100) based on their statistical significance. The least absolute shrinkage and selection operator (LASSO) with five-fold cross-validation was subsequently applied to select the optimal subset of features to enhance sparsity and interpretability. Furthermore, clinical features including demographic, pathological and imaging-derived variables were integrated with the final selected radiomics features to develop a comprehensive clinical-radiomics model, aiming to exploit the complementary strengths of both data types for enhanced breast cancer subtype classification.\u003c/p\u003e\n\n\u003cp\u003e\u003cem\u003eMachine learning model development. \u003c/em\u003eTo address the class imbalance inherent in the dataset, the Synthetic Minority Over-sampling Technique from the \u0026ldquo;imbalanced-learn\u0026rdquo; library (version 0.12.4; https://imbalanced-learn.org/) was applied only within the training folds of each cross-validation split (26). Due to the substantial variability in model performance introduced by different machine learning algorithms, a comprehensive set of machine learning classifiers was subsequently evaluated on the combined clinical-radiomics features, including logistic regression (27), k-nearest neighbors (KNN) (28), naive Bayes (29), support vector machine (SVM) (30), decision tree (31), random forest (32), Adaptive Boosting (AdaBoost) (33), gradient boosting decision tree (GBDT) (34) and Extreme Gradient Boosting (XGBoost) (35). Model development was performed using five-fold cross-validation, with the best-performing fold selected for each model based on validation metrics. All models were trained using default hyperparameters (Supplement Table 2). The models chosen throughout this procedure were subsequently externally validated on independent cohorts from Centers 2 and 3 to assess their generalizability and predictive performance.\u003c/p\u003e\n\n\u003cp\u003e\u003cem\u003eModel evaluation and statistical analysis\u003c/em\u003e. The performance of the radiomics models was comprehensively evaluated using several metrics, including the area under the curve (AUC), sensitivity, specificity and accuracy. Receiver operating characteristic (ROC) analysis was performed to assess the discriminatory ability of each model. To facilitate model interpretability, SHapley Additive exPlanations (SHAP) values were calculated to quantify the contribution of each feature to model predictions. Comparative analysis of different classifiers was performed using DeLong\u0026rsquo;s test to determine the statistical significance of differences in AUCs. Clinical utility of each model was further assessed using decision curve analysis (DCA) by calculating the net benefit across a range of threshold probabilities. To ensure the reliability of the research findings, bootstrapping with 1,000 iterations was employed to estimate confidence intervals (CI) for performance metrics. For continuous variables, normality was assessed using the Shapiro\u0026ndash;Wilk test implemented in Python (SciPy library, version 1.31.1) (Supplement Table S3). Group comparisons were performed using unpaired Student\u0026rsquo;s t-test when the data conformed to a normal distribution, and the non-parametric Mann-Whitney U test was applied when normality was not met. Categorical variables were analyzed using the Fisher\u0026rsquo;s exact test. P\u0026lt;0.05 was considered to indicate a statistically significant difference.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cem\u003ePatient characteristics\u003c/em\u003e. The clinical and molecular characteristics of patients with luminal A and luminal B breast cancer across the three independent centers are summarized in Table 1. In the cohort from Center 1, patients with luminal A were significantly older compared with those with the luminal B subtype (mean age, 57.05 vs. 49.00 years; P=0.008). ER and PR levels were also significantly higher in the luminal A group (ER, 79.73 vs. 67.00%; P=0.025; PR, 66.05 vs. 27.07%; P\u0026lt;0.001), whereas Ki-67 expression was significantly lower (16.58 vs. 34.31%; P\u0026lt;0.001) compared with the luminal B group. No significant differences were observed in tumor margins, morphology, HER-2 status or lymph node involvement (all P\u0026gt;0.05). In Center 2, Ki-67 was the only variable showing a statistically significant difference between subtypes, being markedly lower in luminal A tumors compared with luminal B tumors (6.86 vs. 35.00%; P\u0026lt;0.001). Other clinical and pathological variables, including age, ER and PR expression, tumor margins, morphology, HER-2 status and lymph node status, did not differ significantly between luminal A and B subtypes (all P\u0026gt;0.05). Data from Center 3 also revealed significant differences in PR and ER levels (PR, 76.67 vs. 28.75%; P＜0.001; ER, 84.44 vs. 51.67%; P=0.018) in the luminal A group compared with the luminal B group. HER-2 status distribution also varied significantly between the subtypes (P=0.027). Other characteristics, including age, Ki-67, tumor margins, morphology and lymph node status, did not show significant differences (all P\u0026gt;0.05).\u003c/p\u003e\n\n\u003cp\u003e\u003cem\u003eFeature extraction and selection\u003c/em\u003e. In the present study, 1,130 radiomic features were initially extracted from each imaging modality. After dimensionality reduction and feature selection, 24 features from DCE-MRI, 14 from T2WI and 16 from DWI were retained for constructing the respective single-parameter models. For the multiparametric fusion models, 14 features were selected for combining DCE-MRI and T2WI, 11 for DCE-MRI and DWI, 13 for T2WI and DWI and 14 for the integrative model incorporating all three MRI sequences. In addition, a clinical-radiomics fusion model was developed by integrating 10 clinical variables including age, sex, tumor size, histological type, tumor margins, tumor shape, ER, PR, lymph node status and Ki-67 index. SHAP-based interpretability analysis of the clinical-radiomics fusion model revealed that clinical and imaging features contributed substantially to subtype classification (Fig. 3, Supplement Fig. 2). Among the top-ranking predictors, Ki-67 index, PR status and patient age demonstrated the most significant influence on the model\u0026rsquo;s output, underscoring the critical role of clinical biomarkers. In parallel, radiomic features derived from DWI (original_shape_MeshVolume) and DCE-MRI (wavelet-LHL_glszm_LargeAreaEmphasis) also ranked highly, indicating that imaging phenotypes provided complementary predictive information. To assess the stability of feature importance across centers, SHAP analysis was extended to both external validation cohorts (Centers 2 and 3). Across all three cohorts, clinical variables (including Ki-67, PR status and age), together with several radiomics features (e.g., DWI_original_shape_MeshVolume), consistently exhibited high importance, indicating that the model relied on stable and generalizable predictors rather than cohort-specific patterns.\u003c/p\u003e\n\n\u003cp\u003e\u003cem\u003eComparative evaluation of machine learning classifiers\u003c/em\u003e. The comparative performance of the machine learning classifiers across the three centers is depicted in Fig. 4. Classifiers such as KNN, naive Bayes, SVM, and logistic regression exhibited consistently suboptimal performance, with AUC values \u0026lt;0.70 in all cohorts, indicating limited discriminative capability. Although the decision tree classifier yielded the highest AUC in Center 1 (0.9450), its performance declined substantially in the external validation cohorts from Center 2 and Center 3 (AUCs of 0.7167 and 0.7083, respectively), indicating limited generalizability. By contrast, the random forest classifier demonstrated consistently robust performance across all cohorts, achieving AUCs of 0.8942, 0.8695 and 0.8935 in Centers 1, 2 and 3, respectively. The GBDT classifier also showed favorable performance in Center 1 and Center 3, with AUCs of 0.9427 and 0.8750, respectively; however, its discriminative ability declined in Center 2 (AUC =0.8128). Among all evaluated classifiers, XGBoost achieved the most stable and high-level performance across the three centers, with AUCs of 0.9392, 0.8966 and 0.8704, respectively. Pairwise AUC comparisons using DeLong\u0026rsquo;s test demonstrated that XGBoost outperformed most of the other classifiers across the three cohorts (Supplement Fig.1), with the majority of comparisons reaching statistical significance (P \u0026lt; 0.05), reflecting superior generalizability and predictive robustness. Consequently, XGBoost was selected as the optimal model for subsequent analyses in the present study.\u003c/p\u003e\n\n\u003cp\u003e\u003cem\u003eClassification performance of the radiomics models. \u003c/em\u003eRadiomics models based on single imaging modalities showed moderate classification performance, while multimodal fusion models markedly improved discrimination between luminal A and luminal B subtypes (Fig. 5). The clinical-radiomics fusion model consistently achieved the highest AUCs of 0.9392, 0.8966 and 0.8704 in Centers 1, 2 and 3, respectively, followed by the fusion of all imaging features (T2WI, DCE-MRI, DWI), which yielded AUCs of 0.9088, 0.8621 and 0.8426. By contrast, the T2WI-based model demonstrated the worst performance across all centers, with AUCs of 0.8018, 0.7340 and 0.7083, respectively.\u003c/p\u003e\n\u003cp\u003eIn addition, the classification metrics across the three centers consistently supported the superior performance of the clinical-radiomics fusion model, as detailed in Tables 2-4 (Center 1-3). In Center 1, the fusion model achieved a sensitivity of 0.8000, specificity of 0.9474, positive predictive value (PPV) of 0.9730 and negative predictive value (NPV) of 0.6667, alongside an AUC of 0.9392. Similar results were observed in Centers 2 and 3, with accuracies of 0.9167 (PPV, 0.9643) and 0.7667 (PPV, 0.6667) and AUCs of 0.8966 and 0.8704, respectively. Notably, including clinical variables markedly enhanced sensitivity and NPV, indicating improved identification of luminal B cases. Among models without clinical data, the fusion of all imaging features yielded the best results across all cohorts (AUCs, 0.9088, 0.8621, 0.8426), substantially outperforming unimodal models. By contrast, single-modality models, particularly those based on T2WI, showed consistently lower sensitivity, specificity and AUC values, with reduced generalizability. These results demonstrate that integrating radiomics features from multiple MRI sequences with clinical variables yields a reliable and generalizable model for subtype classification, supporting its potential applicability in real-world multi-center settings.\u003c/p\u003e\n\u003cp\u003eDCA was performed across Centers 1 to 3 to further assess the predictive models\u0026rsquo; clinical utility, as shown in Fig. 6. In Center 1, the clinical-radiomics model demonstrated the most favorable net benefit when the threshold probability \u0026gt;0.3, followed by the fusion model incorporating all imaging modalities, and the model combining DCE-MRI and DWI features. Models based on single imaging modalities exhibited limited net benefit, especially at higher thresholds. A similar trend was observed in Center 2, where the clinical-radiomics model maintained the most notable net benefit, while multimodal fusion models remained superior to unimodal approaches. In Center 3, although the overall net benefits declined, the clinical-radiomics model continued to outperform others, particularly within the low-to-intermediate threshold range (0.2-0.5). Single-modality models exhibited limited clinical value across most threshold probabilities, highlighting the superior effectiveness of the clinical-radiomics fusion model in enhancing decision-making and providing more reliable clinical guidance.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eDifferences in model performance\u003c/em\u003e. To further evaluate the differences in discriminatory performance among the eight predictive models (Model 1, DCE-based radiomics, Model 2: T2WI-based radiomics, Model 3: DWI-based radiomics, Model 4: T2WI + DCE fusion, Model 5: T2WI + DWI fusion, Model 6: DCE + DWI fusion, Model 7: T2WI + DCE + DWI fusion, and Model 8: T2WI + DCE + DWI + Clinical fusion), pairwise DeLong tests were performed to compare AUCs across the three centers, as illustrated in Fig. 7. In Center 1, most comparisons (Model 1 vs. Models 4, 7, 8; Model 2 vs. Models 4, 6, 7, 8; Model 3 vs. Models 4, 8; and Model 4 vs. Models 5, 6, 7, 8) showed statistically significant differences (P\u0026lt;0.05), especially between single-modality and multi-modality fusion models. The clinical-radiomics fusion model demonstrated significantly higher AUCs compared with most other models. By contrast, differences between it and other multimodal fusion models were not statistically significant (P\u0026gt;0.05), indicating comparable performance among complex models. In Center 2, the clinical-radiomics fusion model consistently outperformed several single-modality and specific dual-modality models, demonstrating its enhanced discriminatory capability. However, the fusion model integrating all imaging modalities without clinical data did not show statistically significant improvements over single-modality models, indicating a diminished incremental benefit. Moreover, no significant differences were observed among the single-modality models, suggesting comparable diagnostic performance within this subgroup. In Center 3, most pairwise comparisons, particularly among fusion-based models, did not demonstrate statistical significance. The lack of discernible differences between the clinical-radiomics fusion model and other fusion approaches suggests a convergence in predictive performance within this cohort, thereby highlighting the inherent challenges in attaining consistent generalizability across heterogeneous patient populations.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eRadiomics quantitatively extracts high-dimensional features (texture, shape, gray-level statistics) from medical images, reflecting tumor microheterogeneity and molecular variations. In the present study, a multiparametric radiomics model was developed by integrating T2WI, DWI and DCE-MRI to differentiate luminal A from luminal B breast cancer subtypes. The fusion model achieved an AUC of 0.9392 in the training cohort and 0.8966 and 0.8704 in two independent test cohorts, significantly outperforming single-modality approaches. This performance was consistently maintained across three institutions despite variations in acquisition protocols, indicating good generalizability and clinical applicability.\u003c/p\u003e\n\u003cp\u003eThe diagnostic advantage of the fusion model can be attributed to the complementary roles of each MRI sequence. DCE-MRI quantifies tumor vascularity and perfusion, both closely associated with angiogenesis. Previous studies have demonstrated that luminal B tumors exhibit higher Ktrans (volume transfer constant) and kep (rate constant) values compared with luminal A tumors, reflecting more active angiogenesis (36, 37). DWI measures water diffusion restriction, with lower apparent diffusion coefficient values in luminal B tumors reflecting higher cellularity (38). T2WI depicts internal architecture and stromal composition, with luminal A lesions more often showing homogeneous signal intensity and abundant fibroglandular stroma (13). By integrating these modalities, the multiparametric clinical-radiomics model captures vascular, microstructural and compositional heterogeneity in a unified framework, likely contributing to its superior performance in Centers 1 and 2 and a robust, albeit slightly lower, performance in Center 3 (39).\u003c/p\u003e\n\u003cp\u003eRecognizing that model generalizability depends not only on feature diversity but also on classifier choice, machine learning algorithms were systematically compared. XGBoost achieved optimal performance across the three centers (AUCs, 0.9392, 0.8966, 0.8704), outperforming traditional classifiers. Pairwise DeLong tests confirmed that the XGBoost-based multiparametric model significantly outperformed most single- and dual-sequence models in Centers 1 and 2. In Center 3, fewer differences reached statistical significance, indicating a narrower performance gap; nevertheless, the multiparametric approach retained high accuracy, highlighting robustness to inter-center variability. In addition, DCA demonstrated that the multiparametric fusion model yielded the highest net benefit across a wide range of threshold probabilities in all three centers, further supporting its potential clinical utility for individualized decision-making. The superior performance of XGBoost is attributable to its ability to model complex, non-linear feature interactions through gradient boosting decision trees. At the same time, built-in regularization mitigates overfitting in high-dimensional data (40-42).\u003c/p\u003e\n\u003cp\u003eAs radiomics often generates thousands of features-numerous redundant or highly associated-the sequential feature selection pipeline (variance thresholding, Pearson correlation filtering, univariate ANOVA and LASSO regression) effectively eliminated non-informative variables while retaining the most discriminative ones (43-45). This strategy yielded a final feature set dominated by clinical biomarkers and shape-related descriptors from DWI, consistent with prior evidence that tumor morphology and proliferation markers carry strong subtype-discriminative value (46-48).\u003c/p\u003e\n\u003cp\u003eThe proposed radiomics framework is designed for two complementary clinical scenarios. First, in pre-biopsy or biopsy-limited settings where Ki-67 and PR are unavailable, the pure imaging-based radiomics model (AUCs of 0.9088, 0.8621, and 0.8426 across three centers) can serve as a non-invasive tool for preliminary subtype stratification. Second, in post-biopsy settings where IHC biomarkers are available, the clinical-radiomics fusion model offers incremental value by capturing whole-tumor heterogeneity that localized biopsy sampling may miss(49). This is particularly relevant given that intratumoral heterogeneity can lead to discordance between biopsy and surgical specimens, potentially resulting in subtype misclassification based on IHC alone(50).\u003c/p\u003e\n\u003cp\u003eThis study has several limitations. First, the sample sizes of the external validation cohorts were relatively limited, particularly the luminal A subgroup in Center 2 (n=7), which may affect statistical power and subgroup-specific performance estimates. Although bootstrap resampling was performed to assess metric stability, future studies with larger and more balanced multicenter cohorts are needed to further validate the proposed model. Second, although standardized spatial resampling and z-score intensity normalization were applied to improve cross-center comparability, explicit harmonization methods such as ComBat were not employed due to the limited sample sizes in some external centers, and variations in scanner hardware and vendor-specific reconstruction algorithms may still influence radiomic feature stability. Third, although SMOTE was strictly applied only within training folds to prevent information leakage, synthetic oversampling in limited cohorts may still amplify noise. Fourth, formal inter- or intra-observer reproducibility analysis for tumor segmentation was not performed; future studies should incorporate multi-reader segmentation with intraclass correlation coefficient analysis. Fifth, as this was a retrospective study, the exact number of patients excluded at each step was not systematically recorded. Finally, the conventional radiomics and machine learning framework may overlook complex patterns that could be captured by deep learning approaches, and reliance on complete multi-sequence data limits applicability when sequences are missing.\u003c/p\u003e\n\u003cp\u003eFuture work should aim to address the limitations of the present study. First, although SMOTE was strictly applied only within training folds to prevent information leakage, synthetic oversampling in limited cohorts may still introduce noise. Expanding the dataset to include more centers, scanner types and acquisition protocols will strengthen feature robustness and improve generalizability. Incorporating scanner-specific harmonization techniques, such as ComBat or deep learning-based normalization, may further reduce inter-scanner variability. Second, the incorporation of radiomics-informed deep learning approaches and spatial/multi-scale features could capture more complex non-linear imaging patterns that conventional radiomics might miss, potentially complementing handcrafted features for enhanced tumor characterization. Finally, extending the model to predict additional clinically relevant endpoints, such as response to neoadjuvant chemotherapy, recurrence risk or metastasis probability, would broaden its clinical utility. These expansions, together with prospective multicenter validation, could facilitate the translation of the proposed model into real-world breast cancer management workflows.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe present study was performed according to the Declaration of Helsinki and was approved by the institutional review boards of all participating institutions. Ethical approval was obtained from: (i) Guiyang Maternal and Child Health Care Hospital/Guiyang Children\u0026apos;s Hospital (approval no. 2024-33; approved April 2024); (ii) Guiyang Second People\u0026apos;s Hospital/The Affiliated Jinyang Hospital of Guizhou Medical University (approval no. 2022-2; approved March 2022); and (iii) Zunyi First People\u0026apos;s Hospital/The Third Affiliated Hospital of Zunyi Medical University (approval no. 2024-1-439; approved August 2024). Due to the retrospective nature of the study and the use of anonymized patient data, the requirement for written informed consent was waived by all institutional review boards. All procedures complied with the Declaration of Helsinki and relevant national and institutional guidelines for research involving human subjects.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data generated in the present study may be requested from the corresponding author.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe present work was supported by the Guizhou Provincial Basic Research Program (Natural Science; grant no. QKHJC-ZK-2023-006) and the Guiyang Health Bureau Science and Technology Project (grant no. ZWJKJ-2021-21).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHZ conceived the study and drafting of the manuscript. DY was responsible for data processing, including analysis and interpretation. XZ designed the study design, critically revised the manuscript. SL, XL and LW contributed to the acquisition of clinical and imaging data. HZ and DY confirm the authenticity of all the raw data. All authors read and approved the final version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eInternational WCRF. Breast cancer statistics London: World Cancer Research Fund International; 2022 [Available from: https://www.wcrf.org/preventing-cancer/cancer-statistics/breast-cancer-statistics.\u003c/li\u003e\n\u003cli\u003ePerou CM, S\u0026oslash;rlie T, Eisen MB, Van De Rijn M, Jeffrey SS, Rees CA, et al. Molecular portraits of human breast tumours. nature. 2000;406(6797):747-52.\u003c/li\u003e\n\u003cli\u003eParker JS, Mullins M, Cheang MC, Leung S, Voduc D, Vickery T, et al. Supervised risk predictor of breast cancer based on intrinsic subtypes. Journal of clinical oncology. 2009;27(8):1160-7.\u003c/li\u003e\n\u003cli\u003eS\u0026oslash;rlie T, Perou CM, Tibshirani R, Aas T, Geisler S, Johnsen H, et al. Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proceedings of the National Academy of Sciences. 2001;98(19):10869-74.\u003c/li\u003e\n\u003cli\u003eGiaquinto AN, Sung H, Miller KD, Kramer JL, Newman LA, Minihan A, et al. Breast cancer statistics, 2022. CA: a cancer journal for clinicians. 2022;72(6):524-41.\u003c/li\u003e\n\u003cli\u003eTorre LA, Islami F, Siegel RL, Ward EM, Jemal A. Global cancer in women: burden and trends. Cancer epidemiology, biomarkers \u0026amp; prevention. 2017;26(4):444-57.\u003c/li\u003e\n\u003cli\u003eColleoni M, Sun Z, Price KN, Karlsson P, Forbes JF, Th\u0026uuml;rlimann B, et al. Annual hazard rates of recurrence for breast cancer during 24 years of follow-up: results from the international breast cancer study group trials I to V. Journal of clinical oncology. 2016;34(9):927-35.\u003c/li\u003e\n\u003cli\u003eBago-Horvath Z, Rudas M, Singer CF, Greil R, Balic M, Lax SF, et al. Predictive Value of Molecular Subtypes in Premenopausal Women with Hormone Receptor\u0026ndash;positive Early Breast Cancer: Results from the ABCSG Trial 5. Clinical Cancer Research. 2020;26(21):5682-8.\u003c/li\u003e\n\u003cli\u003eZhang X, Yang Z, Cui W, Zheng C, Li H, Li Y, et al. Preoperative prediction of axillary sentinel lymph node burden with multiparametric MRI-based radiomics nomogram in early-stage breast cancer. European Radiology. 2021;31:5924-39.\u003c/li\u003e\n\u003cli\u003eZhang Y, Zhou Y, Mao F, Yao R, Sun Q. Ki‐67 index, progesterone receptor expression, histologic grade and tumor size in predicting breast cancer recurrence risk: A consecutive cohort study. Cancer Communications. 2020;40(4):181-93.\u003c/li\u003e\n\u003cli\u003eYuen S, Monzawa S, Yanai S, Matsumoto H, Yata Y, Ichinose Y, et al. The association between MRI findings and breast cancer subtypes: focused on the combination patterns on diffusion-weighted and T2-weighted images. Breast Cancer. 2020;27(5):1029-37.\u003c/li\u003e\n\u003cli\u003eMontemurro F, Martincich L, Sarotto I, Bertotto I, Ponzone R, Cellini L, et al. Relationship between DCE-MRI morphological and functional features and histopathological characteristics of breast cancer. European radiology. 2007;17(6):1490-7.\u003c/li\u003e\n\u003cli\u003eDuong DH, Nguyen QT, Nguyen DT, Tran TL, Duong TG. The association of magnetic resonance imaging features with five molecular subtypes of breast cancer. European Journal of Radiology Open. 2024;13:100585.\u003c/li\u003e\n\u003cli\u003eLambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, Van Stiphout RG, Granton P, et al. Radiomics: extracting more information from medical images using advanced feature analysis. European journal of cancer. 2012;48(4):441-6.\u003c/li\u003e\n\u003cli\u003eLambin P, Van Stiphout RG, Starmans MH, Rios-Velazquez E, Nalbantov G, Aerts HJ, et al. Predicting outcomes in radiation oncology\u0026mdash;multifactorial decision support systems. Nature reviews Clinical oncology. 2013;10(1):27-40.\u003c/li\u003e\n\u003cli\u003eLi C, Song L, Yin J. Intratumoral and peritumoral radiomics based on functional parametric maps from breast DCE‐MRI for prediction of HER‐2 and Ki‐67 status. Journal of Magnetic Resonance Imaging. 2021;54(3):703-14.\u003c/li\u003e\n\u003cli\u003eFan M, Cui Y, You C, Liu L, Gu Y, Peng W, et al. Radiogenomic signatures of oncotype DX recurrence score enable prediction of survival in estrogen receptor\u0026ndash;positive breast cancer: A multicohort study. Radiology. 2022;302(3):516-24.\u003c/li\u003e\n\u003cli\u003eRizzo S, Botta F, Raimondi S, Origgi D, Fanciullo C, Morganti AG, et al. Radiomics: the facts and the challenges of image analysis. European radiology experimental. 2018;2:1-8.\u003c/li\u003e\n\u003cli\u003eXie T, Wang Z, Zhao Q, Bai Q, Zhou X, Gu Y, et al. Machine learning-based analysis of MR multiparametric radiomics for the subtype classification of breast cancer. Frontiers in oncology. 2019;9:505.\u003c/li\u003e\n\u003cli\u003eParmar C, Grossmann P, Bussink J, Lambin P, Aerts HJ. Machine learning methods for quantitative radiomic biomarkers. Scientific reports. 2015;5(1):1-11.\u003c/li\u003e\n\u003cli\u003eSaha A, Harowicz MR, Grimm LJ, Kim CE, Ghate SV, Walsh R, et al. A machine learning approach to radiogenomics of breast cancer: a study of 922 subjects and 529 DCE-MRI features. British journal of cancer. 2018;119(4):508-16.\u003c/li\u003e\n\u003cli\u003eLi H, Zhu Y, Burnside ES, Huang E, Drukker K, Hoadley KA, et al. Quantitative MRI radiomics in the prediction of molecular classifications of breast cancer subtypes in the TCGA/TCIA data set. NPJ breast cancer. 2016;2:16012.\u003c/li\u003e\n\u003cli\u003eJi Y, Whitney HM, Li H, Liu P, Giger ML, Zhang X. Differences in molecular subtype reference standards impact AI-based breast cancer classification with dynamic contrast-enhanced MRI. Radiology. 2023;307(1):e220984.\u003c/li\u003e\n\u003cli\u003eDecoux A, Duron L, Habert P, Roblot V, Arsovic E, Chassagnon G, et al. Comparative performances of machine learning algorithms in radiomics and impacting factors. Scientific Reports. 2023;13(1):14069.\u003c/li\u003e\n\u003cli\u003eGrimm LJ, Johnson KS, Marcom PK, Baker JA, Soo MS. Can breast cancer molecular subtype help to select patients for preoperative MR imaging? Radiology. 2015;274(2):352-8.\u003c/li\u003e\n\u003cli\u003eLema\u0026Atilde;Žtre G, Nogueira F, Aridas CK. Imbalanced-learn: A python toolbox to tackle the curse of imbalanced datasets in machine learning. Journal of machine learning research. 2017;18(17):1-5.\u003c/li\u003e\n\u003cli\u003eHosmer Jr DW, Lemeshow S, Sturdivant RX. Applied logistic regression: John Wiley \u0026amp; Sons; 2013.\u003c/li\u003e\n\u003cli\u003ePeterson LE. K-nearest neighbor. Scholarpedia. 2009;4(2):1883.\u003c/li\u003e\n\u003cli\u003eWebb GI. Na\u0026iuml;ve bayes. Encyclopedia of machine learning and data mining: Springer; 2017. p. 895-6.\u003c/li\u003e\n\u003cli\u003eHearst MA, Dumais ST, Osuna E, Platt J, Scholkopf B. Support vector machines. IEEE Intelligent Systems and their applications. 1998;13(4):18-28.\u003c/li\u003e\n\u003cli\u003eSong Y-Y, Lu Y. Decision tree methods: applications for classification and prediction. Shanghai archives of psychiatry. 2015.\u003c/li\u003e\n\u003cli\u003eBreiman L. Random forests. Machine learning. 2001;45(1):5-32.\u003c/li\u003e\n\u003cli\u003eMargineantu DD, Dietterich TG, editors. Pruning adaptive boosting. ICML; 1997.\u003c/li\u003e\n\u003cli\u003eFriedman JH. Greedy function approximation: a gradient boosting machine. Annals of statistics. 2001:1189-232.\u003c/li\u003e\n\u003cli\u003eChen T, Guestrin C, editors. Xgboost: A scalable tree boosting system. Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining; 2016.\u003c/li\u003e\n\u003cli\u003eHan B, Pan S, Ruan M. Predictive value of neoadjuvant chemotherapy on breast cancer subtypes evaluated by DCE-MRI and MRI perfusion parameters. Indian Journal of Experimental Biology (IJEB). 2023;61(10).\u003c/li\u003e\n\u003cli\u003eLuo H-B, Du M-Y, Liu Y-Y, Wang M, Qing H-M, Wen Z-p, et al. Differentiation between luminal a and B molecular subtypes of breast cancer using pharmacokinetic quantitative parameters with histogram and texture features on preoperative dynamic contrast-enhanced magnetic resonance imaging. Academic Radiology. 2020;27(3):e35-e44.\u003c/li\u003e\n\u003cli\u003eMontemezzi S, Camera L, Giri MG, Pozzetto A, Cali\u0026ograve; A, Meliad\u0026ograve; G, et al. Is there a correlation between 3T multiparametric MRI and molecular subtypes of breast cancer? European Journal of Radiology. 2018;108:120-7.\u003c/li\u003e\n\u003cli\u003eHuang Y, Wang X, Cao Y, Li M, Li L, Chen H, et al. Multiparametric MRI model to predict molecular subtypes of breast cancer using Shapley additive explanations interpretability analysis. Diagnostic and interventional imaging. 2024;105(5):191-205.\u003c/li\u003e\n\u003cli\u003eZhou Y, Lin G, Chen W, Chen Y, Shi C, Peng Z, et al. Multiparametric MRI-based Interpretable Machine Learning Radiomics Model for Distinguishing Between Luminal and Non-luminal Tumors in Breast Cancer: A Multicenter Study. Academic Radiology. 2025.\u003c/li\u003e\n\u003cli\u003eVamvakas A, Tsivaka D, Logothetis A, Vassiou K, Tsougos I. Breast cancer classification on multiparametric MRI\u0026ndash;increased performance of boosting ensemble methods. Technology in cancer research \u0026amp; treatment. 2022;21:15330338221087828.\u003c/li\u003e\n\u003cli\u003eYimit Y, Yasin P, Tuersun A, Wang J, Wang X, Huang C, et al. Multiparametric MRI-based interpretable radiomics machine learning model differentiates medulloblastoma and ependymoma in children: a two-center study. Academic Radiology. 2024;31(8):3384-96.\u003c/li\u003e\n\u003cli\u003eDemircioğlu A. Benchmarking feature selection methods in radiomics. Investigative radiology. 2022;57(7):433-43.\u003c/li\u003e\n\u003cli\u003eZhang W, Guo Y, Jin Q. Radiomics and its feature selection: a review. Symmetry. 2023;15(10):1834.\u003c/li\u003e\n\u003cli\u003ePerniciano A, Loddo A, Di Ruberto C, Pes B. Insights into radiomics: impact of feature selection and classification. Multimedia Tools and Applications. 2024:1-27.\u003c/li\u003e\n\u003cli\u003eHuang T, Fan B, Qiu Y, Zhang R, Wang X, Wang C, et al. Application of DCE-MRI radiomics signature analysis in differentiating molecular subtypes of luminal and non-luminal breast cancer. Frontiers in Medicine. 2023;10:1140514.\u003c/li\u003e\n\u003cli\u003eZhang S, Wang X, Yang Z, Zhu Y, Zhao N, Li Y, et al. Intra-and peritumoral radiomics model based on early DCE-MRI for preoperative prediction of molecular subtypes in invasive ductal breast carcinoma: a multitask machine learning study. Frontiers in Oncology. 2022;12:905551.\u003c/li\u003e\n\u003cli\u003eLeithner D, Mayerhoefer ME, Martinez DF, Jochelson MS, Morris EA, Thakur SB, et al. Non-invasive assessment of breast cancer molecular subtypes with multiparametric magnetic resonance imaging radiomics. Journal of clinical medicine. 2020;9(6):1853.\u003c/li\u003e\n\u003cli\u003eYe D-M, Wang H-T, Yu T. The application of radiomics in breast MRI: a review. Technology in Cancer Research \u0026amp; Treatment. 2020;19:1533033820916191.\u003c/li\u003e\n\u003cli\u003eAhn S, Lee J, Cho M-S, Park S, Sung SH. Evaluation of Ki-67 index in core needle biopsies and matched breast cancer surgical specimens. Archives of pathology \u0026amp; laboratory medicine. 2018;142(3):364-8.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1. Baseline clinicopathological characteristics of patients in Centers 1, 2 and 3.\u0026nbsp;\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"103%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eCharacteristic\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" style=\"width: 28px;\"\u003e\n \u003cp\u003eCenter 1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" style=\"width: 28px;\"\u003e\n \u003cp\u003eCenter 2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" style=\"width: 28px;\"\u003e\n \u003cp\u003eCenter 3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003eLuminal A (n=19)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003eLuminal B (n=45)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003eLuminal A (n=7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003eLuminal B (n=29)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003eLuminal A (n=18)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003eLuminal B (n=12)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eSex\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.654\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eFemale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e19 (100%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e45\u003c/p\u003e\n \u003cp\u003e(100%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003cp\u003e(100%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e29\u003c/p\u003e\n \u003cp\u003e(100%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e16(88.89%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e12(100%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eMale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e2(11.11%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eAge, years\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e57.05\u0026plusmn;10.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e49.00\u0026plusmn;9.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.0076\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e55.00\u0026plusmn;13.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e51.34\u0026plusmn;12.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.5337\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e51.33\u0026plusmn;12.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e54.25\u0026plusmn;12.30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.536\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eHistological type\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.4802\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.640\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e19\u0026nbsp;(100%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e45\u003c/p\u003e\n \u003cp\u003e(100%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e6(85.71%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e27(93.10%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e13(72.22%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e10(83.33%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003cp\u003e(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e1(3.45%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e1(5.56%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e1(14.29%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e1(3.45%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e4(22.22%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e2(16.67%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eTumor margins\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.6540\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e1.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026gt;0.9999\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e12(63.16%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e24(53.33%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e4(57.14%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e16(55.17%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e8(44.44%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e6(50.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e7(36.84%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e21(46.67%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e3(42.86%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e13(44.83%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e10(55.56%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e6(50.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eTumor morphology\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.3961\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.1625\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026gt;0.9999\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e4(21.05%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e16(35.56%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e2(28.57%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e1(3.45%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e5(27.78%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e4(33.33%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e15(78.95%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e29(64.44%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e5(71.43%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e28(96.55%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e13(72.22%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e8(66.67%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eER, %)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e79.73\u0026plusmn;18.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e67.00\u0026plusmn;22.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.0252\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e81.43\u0026plusmn;15.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e78.45\u0026plusmn;22.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.6879\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e84.44\u0026plusmn;13.27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e51.67\u0026plusmn;40.47\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.018\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003ePR, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e66.05\u0026plusmn;18.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e27.07\u0026plusmn;27.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e43.57\u0026plusmn;36.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e47.07\u0026plusmn;38.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.8279\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e76.67\u0026plusmn;12.49\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e28.75\u0026plusmn;30.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eHER-2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.2006\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.7429\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.027\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e7(36.84%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e11(24.44%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e3(42.86%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e10(34.48%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e9(50.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e5(41.67%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e8(42.11%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e20(44.44%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e4(57.14%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e17(58.62%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e9(50.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e3(25.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e4(21.05%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e6(13.33%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e2(6.90%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e8(17.78%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0(0.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e4(33.33%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eKi-67, %)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e16.58\u0026plusmn;6.02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e34.31\u0026plusmn;20.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e6.86\u0026plusmn;3.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e35.00\u0026plusmn;20.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e8.89\u0026plusmn;4.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e12.92\u0026plusmn;10.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.226\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eLymph node status\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.1795\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.4968\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e0.594\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e10 (52.63%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e14(31.11%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e2(28.57%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e15(51.72%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e12(66.67%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e6(50.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e9(47.37%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e31(68.89%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e5(71.43%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e14(48.28%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e6(33.33%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e6(50.00%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eBaseline clinicopathological characteristics of patients in the training and test cohorts. Data are presented as mean \u0026plusmn; standard deviation or as n (%). \u0026ndash; indicates not applicable. P\u003cem\u003e-\u003c/em\u003evalues were calculated using the independent-samples t-test for continuous variables and the \u0026chi;\u003csup\u003e2\u003c/sup\u003e test or Fisher\u0026rsquo;s exact test for categorical variables, as appropriate.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTable 2. Diagnostic performance of different MRI modalities in differentiating luminal A from luminal B breast cancers in Center 1.\u0026nbsp;\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"127%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 9px;\"\u003e\n \u003cp\u003eModality\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003ePPV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 16px;\"\u003e\n \u003cp\u003eNPV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eDCE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.6889 (0.5500, 0.8222)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8947 (0.7333, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7500 (0.6406, 0.8438)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9394 (0.8462, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 16px;\"\u003e\n \u003cp\u003e0.5484 (0.3824, 0.7273)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8082 (0.6910, 0.9103)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eT2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.6889 (0.5555, 0.8182)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8421 (0.6663, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7344 (0.6250, 0.8285)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9118 (0.8107, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 16px;\"\u003e\n \u003cp\u003e0.5333 (0.3571, 0.7200)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8018 (0.6860, 0.9015)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eDWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7111 (0.5743, 0.8409)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8947 (0.7391, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7656 (0.6719, 0.8594)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9412 (0.8571, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 16px;\"\u003e\n \u003cp\u003e0.5667 (0.3714, 0.7500)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8363 (0.7168, 0.9341)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eT2-DCE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.5778 (0.4389, 0.7209)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7895 (0.5833, 0.9524)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.6406 (0.5156, 0.7500)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8667 (0.7308, 0.9677)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 16px;\"\u003e\n \u003cp\u003e0.4412 (0.2666, 0.5947)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7427 (0.6172, 0.8557)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eT2-DWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7556 (0.6250, 0.8810)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8947 (0.7368, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7969 (0.7031, 0.8906)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9444 (0.8610, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 16px;\"\u003e\n \u003cp\u003e0.6071 (0.4090, 0.7778)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8772 (0.7765, 0.9527)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eDCE-DWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7556 (0.6249, 0.8751)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9474 (0.8235, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8125 (0.7188, 0.9062)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9714 (0.9062, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 16px;\"\u003e\n \u003cp\u003e0.6207 (0.4375, 0.7826)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8655 (0.7578, 0.9512)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eT2-DCE-DWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7778 (0.6591, 0.8914)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9474 (0.8235, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8281 (0.7344, 0.9219)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9722 (0.9024, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 16px;\"\u003e\n \u003cp\u003e0.6429 (0.4687, 0.8078)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9088 (0.8313, 0.9758)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eT2-DCE-DWI-clinical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8000 (0.6807, 0.9000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9474 (0.8180, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8438 (0.7500, 0.9219)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9730 (0.9143, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16px;\"\u003e\n \u003cp\u003e0.6667 (0.4848, 0.8400)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9392 (0.8682, 0.9859)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eValues are presented as mean (95% confidence interval). Sensitivity, specificity, PPV and NPV were calculated using the luminal B subtype as the positive class. DCE-MRI; dynamic contrast-enhanced magnetic resonance imaging; T2WI, T2-weighted imaging; DWI, diffusion-weighted imaging; PPV, positive predictive value; NPV, negative predicative value; AUC, area under the curve.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTable 3. Diagnostic performance of different MRI modalities in Center 2.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"127%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003eModality\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003ePPV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003eNPV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003eDCE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8621 (0.7306, 0.9677)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.5714 (0.1667, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8056 (0.6667, 0.9167)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8929 (0.7600, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.5000 (0.1667, 0.8580)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7685 (0.5444, 0.9697)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003eT2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8621 (0.7272, 0.9677)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.5714 (0.1667, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8056 (0.6667, 0.9167)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8929 (0.7692, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.5000 (0.1429, 0.8571)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7340 (0.4285, 0.9609)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003eDWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.6897 (0.5333, 0.8519)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7143 (0.3333, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.6944 (0.5556, 0.8333)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9091 (0.7777, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.3571 (0.1000, 0.6250)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7463 (0.4318, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003eT2-DCE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7931 (0.6428, 0.9286)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.5714 (0.1533, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7500 (0.6111, 0.8889)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8846 (0.7600, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.4000 (0.1111, 0.7143)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7734 (0.4683, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003eT2-DWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7931 (0.6296, 0.9286)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8571 (0.5000, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8056 (0.6667, 0.9167)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9583 (0.8571, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.5000 (0.2000, 0.7857)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8128 (0.4853, 0.9922)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003eDCE-DWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7931 (0.6333, 0.9286)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.5714 (0.1667, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7500 (0.6111, 0.8889)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8846 (0.7600, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.4000 (0.1000, 0.7276)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8399 (0.6617, 0.9798)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003eT2-DCE-DWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8621 (0.7143, 0.9677)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8571 (0.5000, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8611 (0.7500, 0.9722)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9615 (0.8750, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.6000 (0.2727, 0.8889)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8621 (0.5911, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003eT2-DCE-DWI-clinical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9310 (0.8276, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8571 (0.5000, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9167 (0.8056, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9643 (0.8846, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7500 (0.4000, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8966 (0.7059, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eValues are presented as mean (95% confidence interval). Sensitivity, specificity, PPV and NPV were calculated using the luminal B subtype considered as the positive class. DCE-MRI; dynamic contrast-enhanced magnetic resonance imaging; T2WI, T2-weighted imaging; DWI, diffusion-weighted imaging; PPV, positive predictive value; NPV, negative predicative value; AUC, area under the curve.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTable 4. Diagnostic performance of different MRI modalities in Center 3.\u0026nbsp;\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"125%\" class=\"fr-table-selection-hover\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eModality\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003ePPV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 14px;\"\u003e\n \u003cp\u003eNPV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eDCE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.6667 (0.3750, 0.9286)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.5556 (0.3330, 0.7895)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.6000 (0.4333, 0.7667)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.5000 (0.2628, 0.7500)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7143 (0.4615, 0.9286)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7685 (0.5700, 0.9277)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eT2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.6667 (0.3636, 0.9167)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.8333 (0.6471, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7667 (0.6000, 0.9000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7273 (0.4000, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7895 (0.5882, 0.9474)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7083 (0.4976, 0.8929)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eDWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.6667 (0.3846, 0.9093)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.8333 (0.6316, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7667 (0.6000, 0.9000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7273 (0.4286, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7895 (0.5556, 0.9474)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7593 (0.5486, 0.9398)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eT2-DCE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.6667 (0.3844, 0.9167)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7778 (0.5714, 0.9444)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7333 (0.5667, 0.9000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.6667 (0.3636, 0.9231)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7778 (0.5882, 0.9474)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.8009 (0.6294, 0.9500)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eT2-DWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.5833 (0.2857, 0.8571)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7222 (0.5000, 0.9286)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.6667 (0.5000, 0.8333)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.5833 (0.3000, 0.8750)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7222 (0.5000, 0.9286)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7593 (0.5503, 0.9260)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eDCE-DWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7500 (0.4545, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.6667 (0.4284, 0.8824)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7000 (0.5333, 0.8667)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.6000 (0.3635, 0.8462)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8000 (0.5554, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7315 (0.5177, 0.9051)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eT2-DCE-DWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.8333 (0.6000, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7222 (0.5263, 0.9167)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7667 (0.6333, 0.9000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.6667 (0.4286, 0.9091)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8667 (0.6667, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.8426 (0.6832, 0.9778)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eT2-DCE-DWI-clinical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15px;\"\u003e\n \u003cp\u003e0.8333 (0.5714, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7222 (0.5000, 0.9231)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15px;\"\u003e\n \u003cp\u003e0.7667 (0.6325, 0.9000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15px;\"\u003e\n \u003cp\u003e0.6667 (0.4209, 0.9091)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8667 (0.6667, 1.0000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15px;\"\u003e\n \u003cp\u003e0.8704 (0.7083, 0.9861)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eValues are presented as mean (95% confidence interval). Sensitivity, specificity, PPV and NPV were calculated using the luminal B subtype considered as the positive class. DCE-MRI; dynamic contrast-enhanced magnetic resonance imaging; T2WI, T2-weighted imaging; DWI, diffusion-weighted imaging; Clinical: clinical features; PPV, positive predictive value; NPV, negative predicative value; AUC, area under the curve.\u003c/p\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, Magnetic resonance imaging, Radiomics, Machine learning, Luminal subtype, XGBoost, SHAP, Multicenter validation","lastPublishedDoi":"10.21203/rs.3.rs-8793807/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8793807/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAccurate preoperative differentiation of luminal A from luminal B breast cancer is critical for individualized treatment planning, yet current assessment relies on invasive biopsy with inherent sampling limitations. This study aimed to develop and validate a multiparametric MRI-based radiomics model for non-invasive luminal subtype classification across independent institutions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis retrospective multicenter study included 130 patients with pathologically confirmed luminal A (n=44) or luminal B (n=86) invasive ductal carcinoma from three centers. Radiomics features were extracted from T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and dynamic contrast-enhanced MRI (DCE-MRI) using PyRadiomics. A four-stage feature selection pipeline (variance thresholding, Pearson correlation, ANOVA F-test, LASSO regression) was applied for dimensionality reduction. Synthetic Minority Over-sampling Technique (SMOTE) was used within training folds to address class imbalance. Nine machine learning classifiers were systematically compared. The model was developed using Center 1 data (n=64) with five-fold cross-validation and externally validated in two independent cohorts [Center 2 (n=36) and Center 3 (n=30)]. Performance was evaluated using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and decision curve analysis (DCA). Bootstrap resampling (1,000 iterations) assessed metric stability.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAmong nine classifiers, XGBoost demonstrated the best performance. The clinical-radiomics fusion model achieved AUCs of 0.939 (95% CI: 0.868–0.986), 0.897 (95% CI: 0.706–1.000), and 0.870 (95% CI: 0.708–0.986) in Centers 1, 2, and 3, respectively. This model significantly outperformed single-sequence approaches in pairwise DeLong tests (P\u0026lt;0.05) and demonstrated superior net benefit in DCA. SHAP analysis revealed that Ki-67 index, progesterone receptor status, patient age, and radiomics features from DWI and DCE-MRI were the most influential predictors, with consistent importance rankings across all three centers.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe multiparametric MRI radiomics model integrating T2WI, DWI, and DCE-MRI demonstrates robust and generalizable performance for differentiating luminal A from luminal B breast cancer across multiple independent centers. SHAP-based interpretability analysis enhances clinical transparency by identifying consistent predictive features. The model offers potential clinical utility as a non-invasive tool when biopsy is contraindicated and as a supplement to immunohistochemistry by providing whole-tumor heterogeneity assessment.\u003c/p\u003e","manuscriptTitle":"Multiparametric MRI-based radiomics combined with machine learning for preoperative differentiation of luminal A and luminal B breast cancer: a multicenter study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-24 00:44:11","doi":"10.21203/rs.3.rs-8793807/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-05-08T13:10:14+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-30T07:16:51+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"213685809538932841905710212693944197550","date":"2026-04-24T10:32:38+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-22T14:59:37+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-20T05:48:31+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"24879111861184356314966768349280187913","date":"2026-04-15T16:15:19+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"293999148263755194330214000408830909502","date":"2026-04-15T14:19:20+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-15T13:04:06+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-03-23T14:10:18+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-06T08:41:21+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-06T08:34:55+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2026-02-05T07:20:41+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":"229dc28c-8195-4642-8627-f1cd86cb90f2","owner":[],"postedDate":"April 24th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Revision requested","date":"2026-05-08T13:10:14+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-30T07:16:51+00:00","index":71,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[],"tags":[],"updatedAt":"2026-05-08T13:25:03+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-24 00:44:11","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8793807","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8793807","identity":"rs-8793807","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}